# General Input vs. Claude
> Claude is a brilliant assistant, and Cowork brings it to your desktop. General Input turns the same models into governed workflows across 850+ systems.
Claude is one of the most capable AI systems in the world, and in 2026 Anthropic has been pushing it well past the chat box. Claude Cowork puts an agent on your desktop with file access, scheduled tasks, and a plugin marketplace. Claude Managed Agents, now in public beta, adds multi-agent orchestration and webhooks for developers. The Claude 5 models themselves are exceptional at reasoning, writing, and code.
So the honest version of this comparison is not "chat app vs. automation platform" anymore. Anthropic is building automation too. The real difference is architectural: Claude's agent products are one lab's application, running one lab's models, licensed per seat, and centered on a person at a computer. General Input is a neutral automation layer that runs Claude models alongside two dozen others, connects to 850+ business systems, and treats a workflow as a governed, monitored, costed unit of production work that your whole team owns.
One more wrinkle that surprises people: the two products are complementary at the subscription level. General Input can use the Claude subscription you already pay for as the engine behind your General Input automations. You do not have to choose between "keep Claude" and "get automation."
## What Claude does well
- **The models.** Claude's frontier models are among the best available for deep reasoning, long documents, and code. General Input offers them too, and many of our users pick Claude models for their hardest steps.
- **Interactive knowledge work.** Brainstorming, drafting, analyzing a contract, debugging alongside you. For thinking out loud with an AI, Claude is the gold standard.
- **Cowork on the desktop.** File system access, recurring scheduled tasks, and a growing plugin marketplace (Slack, Canva, Figma, Box, Clay) make Cowork genuinely useful for personal productivity.
- **A maturing enterprise story.** Role-based access controls, spend limits, and usage analytics arrived for Cowork in 2026, and Managed Agents is Anthropic's play for production agents.
## How they compare
## Where they differ
### One lab's application vs. a neutral layer
Cowork, plugins, and Managed Agents are excellent, and they are also a strategy: they live inside Anthropic's application, run Anthropic's models, and deepen your dependence on one vendor's roadmap and pricing. That is a fine trade for personal productivity. It is a risky one for business process automation, where a workflow you build today should survive the next model generation, the next pricing change, and the next shift in which lab is ahead.
General Input is model-agnostic by design. The same workflow can run on Claude Opus 5 today, GPT-6 tomorrow, and an open-weights model like Kimi K3 or DeepSeek V4 next quarter, without being rebuilt. When the frontier moves, you move with it.
### An assistant's reach vs. an integration catalog
Cowork's plugin marketplace launched with a handful of well-built integrations, and MCP connectors extend it further if you wire them up. General Input ships 850+ integrations with managed OAuth, one-click connection, and 80,000+ documented operations behind a single control plane. When your workflow needs your CRM, your billing system, your support desk, and your data warehouse in one run, catalog depth is the difference between "works out of the box" and "works after an integration project."
### Governance built for teams
General Input treats every run as an auditable event: full execution history with inputs, outputs, timing, and cost, approval gates that pause a workflow until a person signs off, and workspace roles that control who can view, run, edit, or share each workflow and credential. The Agent Firewall goes further than any assistant can: each credential is scoped to the specific operations a workflow is allowed to call, read-only, write-only, or a single endpoint, and the policy is enforced outside the model's context, so a confused or manipulated model cannot exceed it.
### Your Claude subscription, working overtime
General Input connects the AI subscriptions you already pay for, including Claude, to your General Input workspace. The model runs through your existing plan, so automations execute with no per-token platform markup. If your team already has Claude seats, that spend now covers your automation layer too.
## Pricing
With General Input you pay for what your automations actually use, and every run shows its cost broken down by step. With Claude you pay per person, which is simple for a team of chat users and hard to map to "what did this automated process cost us this month?"
## When to use Claude
Use Claude when a person is in the loop by definition: research, writing, analysis, coding sessions, and desktop tasks where Cowork's file access shines. If your need is a brilliant assistant for individuals, Claude is arguably the best there is, and nothing in General Input replaces that interactive experience.
## When to use General Input
- **The work spans systems.** One run that reads HubSpot, checks Postgres, drafts an email, and waits for a manager's approval before sending. Assistants help you do work; workflows do the work.
- **The work happens without you.** Webhook fires at 3am, workflow runs, humans only hear about it if approval is needed or something fails.
- **More than one person owns it.** Shared credentials with role-based access, workflows the whole team can see, run, and improve, and an audit trail an admin can actually export.
- **You refuse to bet the stack on one lab.** Nine providers, two dozen managed models, your own API keys, your existing subscriptions in General Input, or local models on your own hardware.
- **Compliance is in the room.** Approval gates, per-operation credential scoping, full audit logs, and on-prem deployment for regulated environments.
## FAQs
### Can General Input run Claude models?
Yes. Claude Opus 5, Sonnet 5, and Haiku 4.5 are in the managed model catalog, and you can also connect your own Anthropic API key or your existing Claude subscription in General Input.
### Does Claude Cowork replace a workflow platform?
For personal automation it can cover a lot. For team automation it lacks a shared workflow library, per-run cost attribution, credential scoping enforced outside the model, and deployment options beyond Anthropic's cloud. Those are the gaps General Input fills.
### We already pay for Claude seats. Is General Input redundant?
No, and the two stack well: keep Claude for interactive work, and connect the same subscription to your General Input automations, so the marginal model cost of your workflows is zero.
## Better together
This is the rare comparison where "use both" is the literal architecture. Claude for thinking alongside you during the day. General Input for the governed workflows that run your operations around the clock, quite possibly powered by the same Claude subscription.
---
# General Input vs. Zapier
> Zapier connects more apps than anyone. General Input makes AI the engine, with per-run costs, credential firewalling, model choice, and self-hosting.
Zapier defined the category and it is still the biggest name in it. The catalog now spans roughly 9,000 apps, and the product has grown from "Zaps" into a bundled suite: Agents for AI teammates, Copilot for natural language building, Tables for data, Interfaces for internal tools, Canvas for process mapping, Forms, and Chatbots. For simple app-to-app plumbing with near-zero learning curve, Zapier remains the safe default.
But Zapier's move into AI has exposed the seams in its architecture. AI is priced as a surcharge on a task-based billing model: as of June 2026, AI steps consume tasks at a multiplier depending on the model tier you pick, up to five times the base rate for premium models. The platform is a collection of eight products acquired and built at different times, each with its own concepts to learn. And the fundamentals that matter once automation touches sensitive systems, per-run cost attribution, credential scoping enforced outside the model, and self-hosted deployment, are still missing.
General Input starts from the opposite end. It was built AI-native: you describe the workflow, the platform builds it, and the LLM is the engine rather than a billable add-on step. Every run shows its exact cost. Every credential is scoped to the operations a workflow actually needs. And when compliance requires it, the whole platform deploys on your infrastructure.
## What Zapier does well
- **Catalog breadth.** Roughly 9,000 apps. If an obscure tool has a public API, Zapier probably connects to it. This is the moat, and it is real.
- **Zero learning curve for simple Zaps.** "New row in Sheets, send a Slack message" takes two minutes and never breaks.
- **A mature ecosystem.** Templates, community, agencies, and a decade of documentation. You will never lack a tutorial.
- **The bundled suite.** Tables, Interfaces, and Forms let light internal tools live next to the automations that feed them.
## How they compare
## Where they differ
### The AI tax
Zapier's June 2026 pricing change made the economics of AI automation explicit: AI steps are billed by model tier, at one, three, or five times the task rate, on top of plans that already meter every step. A workflow that summarizes tickets with a premium model burns tasks five times faster than the same workflow moving data. Zapier even added a circuit breaker that pauses a Zap for approval when a single step consumes 75 tasks, which tells you how unpredictable run costs have become.
General Input's billing is the inverse: credits map directly to real usage ($25 buys 4,000 credits), and every execution shows exactly what it cost, broken down by model tokens, integration calls, and compute. You can also connect your existing AI subscriptions directly in General Input and use them in your workflows.
### One platform vs. eight products
Zapier's suite grew by accretion: Zaps, Agents, Copilot, Tables, Interfaces, Forms, Canvas, and Chatbots each have their own mental model, and AI agents live in a separate product with separate behavior from your Zaps. In General Input there is one system. App workflows, deterministic code workflows, and AI agents are three types of the same resource, built in the same chat, sharing the same credentials, permissions, execution history, and cost reporting. Skills let you package instructions once and reuse them across all of them, and Apps let you spin up interactive dashboards and tools from a prompt when a workflow needs a front end.
### Security posture
Zapier stores credentials and passes data through its cloud, full stop. General Input encrypts credentials, keeps them out of the model's context entirely, and adds the Agent Firewall: per-operation scoping so a workflow holding your Stripe key can refund but never create payouts, or read your CRM but never delete from it, enforced at the platform layer where a prompt injection cannot reach it. Every credential use is logged. For regulated teams, the entire platform deploys on-prem, which Zapier does not offer at any price.
### When the work needs judgment
Zapier's sweet spot is still the deterministic hop: event in App A, action in App B. The moment the middle of the workflow requires reading a document, making a decision, or writing something a customer will see, you are gluing AI steps onto a task pipeline. General Input was designed for exactly that middle: agents reason across tools mid-run, approval gates put a human before the irreversible step, and voice dictation and chat make building feel like delegating to a colleague rather than configuring a pipeline.
## Pricing
## When to use Zapier
If your automations are high-volume, deterministic, and app-to-app, and one of your tools is too obscure for anyone else's catalog, Zapier is hard to beat. Teams with hundreds of simple Zaps and no AI reasoning in the critical path have little reason to migrate.
## When to use General Input
- **AI is the point, not a step.** Workflows that read, decide, and write need a platform where the model is the engine and not a 5x surcharge.
- **Finance asks what automation costs.** Exact per-run cost, per-workflow attribution, and credit-based billing you can actually forecast.
- **Security reviews your stack.** Credential isolation from the model, per-operation firewall scoping, exportable audit logs, and on-prem deployment.
- **You want model leverage.** Swap between two dozen models from nine labs as the frontier moves, or plug in the subscriptions you already pay for.
- **One system, not eight.** Workflows, agents, code, skills, and apps in a single workspace with one permission model.
## FAQs
### Can General Input replace all my Zaps?
If a Zap touches one of the 850+ integrations General Input supports, yes, and rebuilding is usually as fast as pasting the Zap's description into the workflow builder. For very long-tail apps, keep those Zaps and let a webhook hand off to General Input for the AI-heavy part.
### Is 850+ integrations enough compared to 9,000?
The catalogs overlap where it matters: CRM, email, messaging, billing, databases, support, project management, and the systems mid-market teams actually run on. Zapier's long tail is real, but most teams' automation footprint fits comfortably inside 500 well-documented integrations, and General Input adds new ones weekly.
### Does General Input have an equivalent to Zapier Agents?
Yes, agents are a first-class workflow type, not a separate product. They share credentials, monitoring, approvals, and cost reporting with every other workflow in your workspace.
## Better together
Plenty of teams run both: Zapier for the long tail of simple triggers, General Input for the workflows where AI judgment, cost visibility, and governance matter. A webhook connects the two in minutes.
---
# General Input vs. n8n
> n8n is the developer's automation power tool. General Input gives the rest of the team that power in plain English, with cost transparency and governance.
n8n has had a remarkable run. A $2.5 billion valuation after its late-2025 Series C, backing from Accel and NVIDIA's venture arm, and a genuinely strong product: a node editor with full visibility into data flowing between steps, native AI agent nodes wired to every major model provider, a free self-hostable Community Edition, and a template library in the thousands. For developer-led teams that think in JSON, n8n is the automation power tool of the moment.
The question n8n never quite answers is who else gets to build. Its AI Workflow Builder can scaffold a workflow from a prompt now, metered by monthly AI credits, but the artifact it produces is still a node graph that someone has to understand, debug, and maintain: webhooks, data mapping, expression syntax, error branches. Your ops lead is not going to open that canvas and own it. And the operational layer around the workflows, cost attribution, approvals, credential governance, remains thin or gated to enterprise tiers.
General Input's bet is different: describe the workflow in plain English, get a reviewable, deployable automation with monitoring, approvals, per-run costs, and credential scoping already attached. Power for developers too, through a CLI and API, but never as a prerequisite.
## What n8n does well
- **Developer control.** Every node, every data transformation, every error branch is visible and editable. The full-code escape hatch means no ceiling.
- **Self-hosting for free.** The Community Edition runs on your own Docker host at zero license cost, which is unbeatable for budget-constrained technical teams.
- **AI agent orchestration.** Native nodes for OpenAI, Anthropic, Google, and open models, with agent and tool-use patterns that developers can compose freely.
- **Community and templates.** Thousands of community workflows, active forums, and an ecosystem that keeps growing with the company's valuation.
## How they compare
## Where they differ
### Who gets to build
This is the structural difference. In n8n, the workflow is a node graph, and the graph is the interface: to build or fix anything you reason about nodes, expressions, and data shapes. General Input's interface is a conversation. "When a lead books a demo, pull their company from HubSpot, enrich it, draft a prep brief, and post it to the AE's Slack an hour before the call" becomes a reviewable workflow in about a minute, and the person who described it can maintain it the same way, by talking, or three times faster by voice. Developers still get depth when they want it: a `geni` CLI, API keys, and deterministic code workflows in TypeScript.
### The operations layer
n8n gives you execution logs. General Input gives you an operations layer: every run recorded with inputs, outputs, timing, and exact cost, workflows and credentials shared under role-based permissions on every plan, approval gates that pause sensitive steps for sign-off, and analytics across the workspace. On n8n, several of those are enterprise-tier features or things you assemble yourself.
### Credential security, enforced outside the model
Both platforms store credentials encrypted. The difference is what happens at run time. General Input never places secrets in the model's context, and the Agent Firewall scopes every credential to the specific operations a workflow is allowed to perform: read-only, write-only, or a single endpoint, org-wide, enforced at the platform layer where prompt injection cannot reach. n8n's credentials are available to whatever the workflow's nodes do with them.
### Total cost of ownership
n8n's free self-hosting is genuinely free only if your engineers' time is. You own the Docker host, upgrades, scaling, queue backends, and debugging when a node behaves differently after an update. n8n Cloud starts around $20 a month but meters workflow executions and AI-builder credits by plan. General Input is managed by default, bills transparent credits ($25 buys 4,000) with exact per-run costs, and for teams that need data on their own metal, offers managed on-prem deployment where the platform team carries the operational load with you.
## When to use n8n
Choose n8n when developers own automation end to end and want maximum control: custom nodes in TypeScript, hand-tuned data transformations, self-hosted infrastructure at zero license cost, and a visual debugger for every execution. For developer-led ETL, event routing, and integration plumbing, it is excellent.
## When to use General Input
- **Automation belongs to the whole team.** Ops, finance, support, and marketing build and maintain their own workflows in plain English, with no node graphs in sight.
- **You need to answer "what did this cost?"** Per-run cost breakdowns and workspace analytics, not a self-assembled observability stack.
- **Sensitive systems are involved.** Approval gates, per-operation credential firewalling, and audit logs on standard plans.
- **You want managed, with an exit.** A managed platform day to day, on-prem deployment when compliance demands it, and local models when data cannot leave the building.
- **AI is the workload.** Two dozen managed models from nine labs, auto-routing, effort control, and agents as a first-class workflow type rather than a node pattern.
## FAQs
### Can General Input do the complex branching n8n handles?
Yes, differently. Where n8n expresses complexity visually, General Input expresses it in the workflow definition and in deterministic code steps written in TypeScript. You describe the logic; the platform writes and maintains the code, and you can review every line.
### We already self-host n8n. Why consider General Input?
Keep n8n for the developer plumbing it does well. Teams usually add General Input when non-developers start queuing automation requests, when AI steps need cost attribution, or when a security review asks how credentials are isolated from the model. A webhook bridges the two platforms cleanly.
### Is General Input open source?
No. The platform is commercial with a free tier. What it offers instead on the sovereignty axis: on-prem deployment, bring-your-own keys, and local model support in General Input, so your data and your model spend stay under your control.
## Better together
n8n for developer-owned pipelines and the long tail of hand-tuned integrations. General Input for the AI-native workflows the rest of the team owns, with the governance layer production automation needs. Many teams run exactly that split.
---
# General Input vs. Composio
> Composio gives engineers tool-calling infrastructure for agents they code. General Input is the full platform where the whole team builds and governs AI work.
Composio has grown into the default answer to one specific question: "my agent needs to call real tools, how do I skip building fifty OAuth flows?" Its SDKs plug pre-authenticated connectors into LangChain, CrewAI, and custom agent code; its catalog has expanded past 1,000 toolkits; and Rube, its universal MCP server, lets any MCP-capable AI client act across 500+ apps with one install. Nearly 30,000 GitHub stars say developers like the answer.
But Composio is infrastructure, deliberately. It handles auth, tool routing, and API surface, and leaves everything else to the engineer: where the agent runs, how it is scheduled and triggered, what it logs, what it costs, who can change it, and what happens when it needs a human sign-off. That is the right shape for a team shipping a custom agent product. It is the wrong shape for a company that wants its operations automated, because most of the people with automation needs do not write Python.
General Input covers the whole loop: the same managed-auth integration layer (850+ apps, 80,000+ documented operations), plus the platform around it, natural language building, triggers and schedules, execution history with per-run costs, approval gates, credential firewalling, and workspace roles. Developers get a CLI and API keys. Everyone else gets to build without them.
## What Composio does well
- **Tool-calling infrastructure.** Pre-authenticated connectors, clean SDK ergonomics in TypeScript and Python, and managed OAuth that saves real engineering weeks.
- **Rube and MCP leadership.** A universal MCP server that makes 500+ apps available to any MCP-capable client is a genuinely useful primitive, and Composio is ahead here.
- **Catalog scale.** Over 1,000 toolkits and counting, with SOC 2 compliant auth handling underneath.
- **Framework fit.** If your team already lives in LangChain or CrewAI, Composio drops in with minimal ceremony.
## How they compare
## Where they differ
### A layer vs. a loop
Composio hands your code a set of authenticated tools and steps back. Everything that makes automation production-grade remains your build: scheduling, retries, observability, cost accounting, permissions, approvals. General Input closes that loop in the platform. A workflow gets a trigger (webhook, schedule, email, or app event), an execution history with inputs, outputs, timing, and exact cost, approval gates before sensitive actions, and role-based sharing, all before you write a line of code, because you do not have to write one.
### Who files the ticket
With Composio, every new automation is an engineering task by definition: someone writes agent code, wires tools, deploys it somewhere, and owns it forever. With General Input, the account manager who wants "summarize every new support escalation and post it to my channel with a suggested reply" builds it themselves by describing it, and engineering only gets involved for the workflows that genuinely need custom code, through deterministic TypeScript steps, the `geni` CLI, and API keys.
### Security that does not depend on the agent behaving
Composio's auth handling is solid, but once a tool is granted to your agent, your code decides what gets called. General Input assumes the model cannot be fully trusted: credentials never enter model context, and the Agent Firewall restricts each credential to the operations that workflow is allowed, read-only, write-only, or a single endpoint, with the policy enforced at the platform layer and every credential use logged. A prompt-injected agent cannot call what the firewall never exposes.
### Cost visibility
Composio meters API calls (free to 20,000 a month, then paid tiers), but your real cost is spread across it, your model provider, and your hosting. Nobody can answer "what does this automation cost per run?" General Input answers it natively: every execution shows its cost broken down by model tokens, integration calls, and compute, priced in transparent credits ($25 buys 4,000). You can also connect your existing AI subscriptions directly in General Input.
## When to use Composio
Use Composio when you are building an agent product in code and need tool access fast: a customer-facing AI feature, a custom internal agent framework, or MCP connectivity for AI clients your team already uses. Its SDKs and Rube are best-in-class for that job.
## When to use General Input
- **The goal is automated operations, not an agent codebase.** Workflows built by describing them, deployed with triggers, monitored with per-run costs.
- **Non-engineers have automation needs.** The whole team builds; engineering reviews the sensitive ones via approval gates instead of owning every build.
- **Governance is a requirement.** Credential firewalling, audit logs, workspace roles, and on-prem deployment for regulated environments.
- **You want the model layer handled.** Two dozen managed models from nine labs, effort control, auto-routing, your own keys, or your own subscriptions connected in General Input.
- **One bill, one dashboard.** Execution, integrations, and models accounted for in one place instead of three vendors and a hosting platform.
## FAQs
### Is General Input built on Composio?
No. General Input's 850+ integrations, managed OAuth, and 80,000+ documented operations are first-party, maintained in-house with per-operation documentation the agent uses at run time.
### Can my engineers still write code in General Input?
Yes. Deterministic code workflows are TypeScript, reviewable line by line, and the `geni` CLI plus API keys support scripting, CI, and programmatic workflow management. The difference is that code is an option, not the entry ticket.
### Does General Input support MCP?
Not as a gateway product. If exposing tools to third-party MCP clients is your core need, Composio's Rube is the better fit; General Input focuses on running governed automations inside its own platform.
## Better together
If your engineering team ships a custom agent with Composio's SDK, run it side by side with General Input: the custom product handles its niche, and General Input handles the operational automation for the rest of the company, with the monitoring, approvals, and cost accounting neither team wants to rebuild.
---
# General Input vs. Tasklet
> Tasklet shuts down in 2026: free accounts are deleted August 15 and paid accounts September 14. How it compares and how to move your agents to General Input.
Tasklet deserved its fans. It took the "just tell it what to do" approach further than most: describe a task in conversation and its cloud agent figured out the tools, the steps, and the execution order on its own, running in isolated sandboxes with triggers, subagents, and even browser automation for apps without APIs. For quick personal automations with zero upfront setup, it was impressively fast.
That chapter is ending. Tasklet has stopped accepting new signups and announced a wind-down: **free accounts are deleted on August 15, 2026, and paid accounts on September 14, 2026.** If your automations live there, they need a new home in the next few weeks, and anything you have not exported will go with the servers.
Tasklet free accounts are deleted August 15, 2026; paid accounts September 14,
2026. Export your task definitions, trigger configurations, and
connected-account lists before your date. After deletion there is nothing to
recover.
General Input is the closest philosophical successor: you describe work in plain English, the platform builds it, and it runs in the cloud on triggers and schedules. The difference is that General Input was built for the parts Tasklet deferred: a reviewable workflow you can inspect before it runs, approval gates, per-run cost transparency, model choice, team permissions, and a company behind it operating a durable platform.
## What Tasklet got right
- **Conversation as the interface.** No canvas, no nodes. You talked, it worked. General Input shares this DNA, down to voice dictation for building workflows by speaking.
- **Always-on cloud execution.** Agents ran server-side with schedules, webhooks, and email triggers, no laptop required.
- **Sandboxed runs.** Each task executed in an isolated cloud sandbox, a real security decision more platforms should have made.
- **Browser automation.** Driving a real browser for apps without APIs was genuinely useful, and remains the one Tasklet feature without a General Input equivalent today.
## How they compare
## Migrating from Tasklet to General Input
Most Tasklet automations translate directly, because the two platforms share the same core idea. A practical path:
1. **Inventory your tasks.** In Tasklet, list every task, its trigger (schedule, webhook, email), the connected accounts it uses, and what it produces. Copy each task's instruction text; that text is most of your migration.
2. **Recreate by describing.** Paste each task description into General Input's builder chat, or dictate it. The builder produces a structured workflow you can review, edit, and approve before it ever runs, which is the step Tasklet skipped.
3. **Reconnect your apps.** General Input has managed OAuth for 850+ integrations, so the accounts Tasklet reached through its connectors reconnect in a couple of clicks each, with credentials encrypted, isolated from the model, and scoped by the Agent Firewall to only the operations each workflow needs.
4. **Re-point your triggers.** Schedules are configured in the workflow; webhook triggers get a fresh General Input URL to paste into the source system; email-driven tasks watch the Gmail or Outlook inbox you connect.
5. **Add the guardrails Tasklet lacked.** Put approval gates in front of anything that emails a customer or spends money, and check the per-run cost breakdown after the first few executions.
The one gap to plan around: workflows that relied on Tasklet's browser automation. General Input does not drive a browser today, so those flows should move to API-based integrations, which are more reliable at production volume anyway. If the app in question is among the 850+ in the catalog, the API path usually covers what the browser was doing.
Migrating more than a handful of tasks? Book a call and we will walk your
Tasklet inventory with you and rebuild the first workflows together, free.
## Pricing after the move
Tasklet billed credits with no published per-task rate, no rollover, and consumption that varied with task complexity and the intelligence tier its agent chose. General Input's credits are transparent: $25 buys 4,000, every execution reports its exact cost broken down by step, and there is a free plan to migrate onto before committing. If you already pay for ChatGPT, Claude, Gemini, or Grok, you can connect those subscriptions directly in General Input and use them in your workflows.
## FAQs
### Can General Input import my Tasklet tasks automatically?
There is no one-click importer, because Tasklet tasks are free-form agent instructions rather than a portable format. In practice the instruction text is the task: paste it into the builder and review what comes out. Most tasks rebuild in minutes.
### What happens to my Tasklet data after the deadlines?
Per Tasklet's wind-down notice, accounts and their data are deleted on the dates above. Export everything you might need first: instruction text, trigger settings, run history you want to keep, and the list of connected accounts.
### Is General Input going to disappear too?
No platform can promise forever, but the risk profile is different: General Input is a revenue-funded platform business, and your exit options are stronger, workflows are inspectable and exportable, an Enterprise deployment can run on your own infrastructure, and you can connect AI subscriptions you control.
### I used Tasklet's subagents. What is the equivalent?
Agents in General Input can be composed: a workflow can invoke other workflows and skills, so a "research subagent plus writing subagent" pattern becomes two workflows or skills orchestrated by a third.
## The bottom line
Tasklet proved that conversational automation works. General Input is where that idea grows up: the same describe-it-and-it-runs experience, plus the review step, approvals, cost transparency, model freedom, and team governance that production automation needs, from a platform that will still be here next year. Migrate before your deadline, and bring the instruction text; it is most of the work.
---
# General Input vs. OpenAI
> OpenAI builds the frontier models. General Input builds the stable automation layer around them, with 850+ integrations and no single-lab lock-in.
OpenAI makes some of the best models in the world, and its agent platform ambitions are serious: AgentKit brought a visual Agent Builder, a Connector Registry, and ChatKit for embedding agents; Workspace Agents succeeded custom GPTs for enterprises and plug into Slack and Salesforce; and Frontier, launched in February 2026, is the enterprise platform for managing "AI coworkers" with shared business context, permissions, and evaluation.
But 2026 also showed what building on a lab's platform means. In June, OpenAI announced it is winding down Agent Builder and its Evals product, with both gone from the platform on November 30, 2026, barely a year after launch, with teams pointed toward Frontier and code-level SDKs instead. That is the pattern with lab platforms: the tools orbit the lab's strategy, and the lab's strategy moves fast. If your revenue operations run on a builder that gets deprecated, "migrate by November" becomes your problem.
General Input is the counter-position: a neutral automation platform where OpenAI's models are one excellent option among two dozen, where workflows survive vendor pivots because they are not built inside any vendor's application, and where the operational layer, 850+ integrations, triggers, approvals, per-run costs, credential firewalling, is the product rather than a side quest.
## What OpenAI does well
- **The models.** The GPT-6 family is exceptional, and General Input customers use these models heavily. This comparison is about the platform around them, not the models.
- **Developer primitives.** The Agents SDK, Guardrails, and ChatKit are strong building blocks for engineering teams shipping AI products.
- **Enterprise ambition.** Frontier's shared business context, execution environments, and permissions target real enterprise needs, for organizations ready to commit at that level.
- **ChatGPT distribution.** Workspace Agents meet hundreds of millions of users inside a product they already know.
## How they compare
## Where they differ
### Platform stability vs. lab velocity
OpenAI ships fast and prunes fast. Custom GPTs gave way to Workspace Agents; Agent Builder and Evals launched in October 2025 and were marked for shutdown by June 2026; Frontier is the new center of gravity, until the next reorganization of the roadmap. None of this is a criticism of the models, but automation infrastructure has a different job than a research lab's product surface: your invoice-processing workflow should not need a migration plan every fiscal year. General Input's product is the automation layer itself, and workflows built two years ago still run.
### One lab vs. the field
Model leadership has changed hands repeatedly, and pricing moves with every release cycle. A workflow built inside OpenAI's platform runs OpenAI models, permanently. The same workflow in General Input picks the best tool per job: GPT-6 for one step, Claude Opus 5 for deep analysis, Gemini for long context, an open-weights model like Kimi K3 or DeepSeek V4 where cost dominates, with auto-routing and per-step effort control if you would rather not choose by hand. When the leaderboard flips, you flip a setting, not your architecture.
### The integration surface
OpenAI's Connector Registry covers the marquee names. General Input maintains 850+ integrations with 80,000+ documented operations behind one control plane, with managed OAuth and per-operation docs the agent reads at run time. Business automation lives and dies on the long tail: the regional accounting tool, the recruiting ATS, the industry-specific CRM. That tail is the catalog's job, and it is where lab platforms are thinnest.
### Cost you can attribute
OpenAI's dashboard tells you what your org spent on tokens. It cannot tell you what the lead-enrichment workflow cost per lead last month. General Input meters every execution, broken down by model tokens, integration calls, and compute, priced in transparent credits ($25 buys 4,000). And if your team already pays for ChatGPT, you can connect that subscription to your workflows in General Input, so the models you are already buying power the automation layer at no additional per-token cost.
## When to use OpenAI
Build directly on OpenAI when you are shipping an AI product: engineers, code, SDKs, evals you run yourself, and a workload tuned tightly to GPT-6's strengths. If your company is enterprise-scale and all-in on OpenAI, Frontier is the intended home for that commitment.
## When to use General Input
- **You are automating operations, not shipping an AI product.** Workflows connected to your CRM, billing, support, and data systems, built by describing them.
- **Deprecation risk is unacceptable.** Your automations should outlive any lab's quarterly platform strategy.
- **You want the whole model market.** Two dozen managed models from nine labs, plus your own keys, your own subscriptions, or local models on your own hardware.
- **Finance and security have questions.** Per-run cost attribution, approval gates, credential firewalling, audit logs, and on-prem deployment.
- **Non-engineers build too.** Plain-English and voice building for the team, CLI and API keys for the developers.
## FAQs
### Does General Input support GPT-6?
Yes. The GPT-6 family (Astra, Sol, Luna) is in the managed catalog, and you can bring your own OpenAI API key or connect your ChatGPT subscription in General Input.
### We built agents in Agent Builder. What now?
Before the November 30, 2026 shutdown, rebuild the automations you rely on somewhere durable. Most Agent Builder flows translate to General Input by describing the same logic in the builder chat, and the result gains triggers, cost tracking, approvals, and integration reach in the process.
### Is General Input's AI worse because it is not a lab?
General Input does not train models; it runs the labs' best ones. Your workflows use the same GPT-6, Claude 5, and Gemini 3 quality, with the freedom to mix them per step.
## Better together
Most General Input customers are OpenAI customers too: GPT-6 powering workflow steps, chosen because it earned the job, inside a platform that also connects the other 500 systems, prices every run, and will not sunset your automation layer with a blog post.
---
# General Input vs. Make
> Make gives power users a visual canvas for complex data flows. General Input replaces the canvas with a conversation, plus cost transparency and approvals.
Make is the visual automation tool for power users, and it has been busy. AI Agents went into open beta in early 2026 with Maia, Make's AI builder, behind them. The Grid, a visual map of your whole automation landscape, rolled out to all paid users. A Reasoning Panel now traces agent decisions live during a run, and Autopilot gives agents their own cloud computers for work beyond API reach. Meanwhile the pricing model has been shifting from operations to credits, with AI modules consuming more per action than plain data moves.
The canvas remains both the appeal and the tax. Building a scenario still means understanding modules, routers, iterators, aggregators, data structures, and Make's execution model, and the new AI features are layered onto that same architecture. Your marketing manager is still not going to open the canvas and ship a competitor-monitoring workflow, and with credit consumption varying by module type, predicting what a scenario costs has gotten harder, not easier.
General Input takes the opposite path: no canvas to learn, because you describe the workflow and review what the platform builds. AI is the engine rather than a premium module class. And the run-time layer, per-execution cost breakdowns, approval gates, credential firewalling, workspace roles, comes standard rather than assembled.
## What Make does well
- **Visual expressiveness.** For complex data routing, branching, array iteration, and error handling across dozens of modules, Make's canvas is one of the most capable tools anywhere.
- **Catalog depth.** 3,000+ apps with granular control over every field mapping and transformation.
- **The Grid and Reasoning Panel.** Seeing your whole automation landscape, and watching an agent's decisions stream live, are genuinely good observability ideas.
- **Approachable entry pricing.** Paid plans start around $9 a month, cheap for hobbyists and small experiments.
## How they compare
## Where they differ
### Describing vs. diagramming
"When a deal closes in HubSpot, enrich the company, draft an onboarding email, and wait for the account manager to approve before sending" is a working General Input workflow in about a minute, built by typing or speaking that sentence. In Make, it is eight-plus modules to select, connect, and configure, and the person who builds it becomes its permanent maintainer. Make's Maia narrows this gap by assisting inside the canvas, but the artifact is still a canvas, and the canvas still has a learning curve that decides who in your company gets to automate.
### What a run costs
Make's shift toward credit billing made AI's price variable: AI modules consume more credits than standard operations, consumption depends on what the agent does, and unused allowances only roll forward a month. Estimating a scenario's monthly cost means modeling module mix and agent behavior. General Input meters the actual run: every execution reports its exact cost, broken down by model tokens, integration calls, and compute, in credits with a fixed, published rate ($25 buys 4,000). Forecasting is arithmetic, not simulation. You can also connect the AI subscriptions you already pay for directly in General Input.
### Governance and the security review
Make offers scenario permissions and logs; the deeper controls production automation needs are thinner. General Input ships them as the default: approval gates that pause a run until a person signs off, credentials encrypted and never placed in model context, Agent Firewall scoping so each credential can only perform the operations its workflow needs (read-only, write-only, or a single endpoint), every credential use logged, and on-prem deployment when data cannot live in a vendor's cloud. Make is cloud-only.
### Where the AI actually lives
Make added AI to an architecture built for deterministic data routing: agents are modules inside scenarios. General Input was built the other way around: the model is the engine, and determinism is what you add where you need it, via code workflows that run identically every time. Both philosophies work; they optimize for different centers of gravity. If your automations are mostly data plumbing with occasional AI, Make's shape fits. If reasoning is the core of the work, summarizing, deciding, drafting, enriching, the AI-native shape wins.
## When to use Make
Choose Make when a technically minded operations team wants visual control over intricate data flows: heavy branching, array manipulation, and error handling across a deep catalog, at low entry cost. For visually orchestrated ETL between SaaS tools, the canvas earns its keep.
## When to use General Input
- **Everyone builds, not just the canvas experts.** Plain-English and voice building puts automation in the hands of ops, finance, support, and marketing.
- **AI is the workload.** Agents, model choice across nine labs, effort control, and auto-routing, without premium-module credit math.
- **You must answer for costs.** Exact per-run breakdowns and workspace analytics, priced at a published credit rate.
- **Security signs off.** Approval gates, credential firewalling, audit logs, and on-prem deployment.
- **One platform, four building blocks.** App workflows, deterministic code, AI agents, and interactive Apps, sharing credentials, permissions, and monitoring.
## FAQs
### Can General Input handle Make-level complexity without a canvas?
Yes. Complex logic lives in the workflow definition and, where determinism matters, in TypeScript code steps you can review line by line. You express the complexity by describing it; the platform writes and maintains the implementation.
### We have hundreds of Make scenarios. Is migration realistic?
Migrate incrementally. Leave stable data-plumbing scenarios in Make, and move the AI-heavy and approval-needing workflows to General Input first, where the payoff is largest. A webhook connects the two platforms for hybrid flows.
### Does General Input have anything like the Reasoning Panel?
Every run has a full execution trace: inputs, outputs, tool calls, timing, and cost per step, kept as reviewable history rather than a live-only view, and exportable for audits.
## Better together
Make's canvas is a fine home for intricate, developer-adjacent data routing. General Input is the layer where AI does the thinking and the whole team does the building, with the cost and governance story a business can stand behind. Many teams run both, connected by a webhook.
---
# General Input vs. Activepieces
> Activepieces is the leading open-source automation platform. General Input is the managed, AI-native alternative, with cost transparency and approvals.
Activepieces has earned its place as the open-source automation platform to beat. The visual builder is clean, the catalog has grown past 450 integrations through an active community contributing TypeScript pieces, and the platform leaned into AI early: agents can now live inside flows as steps that reason and call tools, human-in-the-loop steps are built in, and its MCP support, both exposing and consuming servers, has made it arguably the largest open-source MCP toolkit anywhere. You can run all of it on your own Docker host, air-gapped if you need to, at zero license cost.
The honest gap is everything around the flows. Self-hosting means you own the infrastructure: upgrades, scaling, backups, and debugging when a community piece misbehaves. The builder, clean as it is, is still a visual flow editor that someone has to learn, and the run-time layer stops at logs: no per-run cost accounting, no credential scoping enforced outside the model, and team governance that thickens only in paid editions.
General Input plays a different position: a managed, AI-native platform where workflows are built by describing them, every run is priced to the cent, credentials are firewalled per operation, and the on-prem option comes with the platform team carrying the operational load alongside you.
## What Activepieces does well
- **Real open source.** The code is public, the community contributes pieces in TypeScript, and you can audit, fork, and extend everything. That transparency is a genuine trust advantage.
- **Self-hosting freedom.** Docker up, air-gap if required, pay nothing per execution. For technical teams with spare ops capacity, the economics are unbeatable.
- **MCP leadership in open source.** Exposing flows as MCP servers and consuming external ones makes Activepieces a natural hub for MCP-centric stacks.
- **AI in the flow.** Agent steps, AI utilities, and human-in-the-loop approvals are integrated rather than bolted on.
## How they compare
## Where they differ
### Who runs the infrastructure
Activepieces self-hosting is genuinely free and genuinely yours, both the control and the pager. Postgres, Redis, workers, upgrades, and piece-level debugging are your team's job forever. General Input is managed by default; when compliance requires your infrastructure, the Enterprise on-prem deployment comes with the platform team handling setup and updates with you. The question is not whether you can run Activepieces, it is whether running it is the best use of the engineers who would.
### Who gets to build
Activepieces' visual builder is one of the friendlier ones, but a flow editor is still a skill: triggers, pieces, branches, data mapping. General Input removes the skill requirement: describe the workflow in plain English, or dictate it, review the structured workflow the builder produces, and deploy. The ops lead who could describe the process is the person who ships it. Developers keep their depth through deterministic TypeScript code steps, the `geni` CLI, and API keys.
### The run-time layer
This is where the two platforms diverge hardest. General Input treats every execution as a governed event: exact cost per run broken down by model tokens, integration calls, and compute; approval gates before sensitive steps; credentials encrypted, kept out of model context, and scoped by the Agent Firewall to only the operations each workflow needs; every credential use logged. Activepieces gives you flow logs and (in paid tiers) project roles, and leaves cost attribution and credential-to-AI isolation for you to reason about.
### The AI economics
Running AI steps in self-hosted Activepieces means bringing your own model keys and watching that spend in your provider's dashboard, disconnected from flow context. General Input connects the two: the run that spent the tokens shows the tokens, priced, per step. You can also connect the ChatGPT, Claude, Gemini, or Grok subscriptions you already pay for directly in General Input and use them in your workflows.
## When to use Activepieces
Choose Activepieces when open source is the requirement: you want auditable code, community pieces you can fork, air-gapped self-hosting at zero license cost, or MCP-first architecture, and you have the engineering capacity to own the deployment. It is the best open-source option in this category.
## When to use General Input
- **Managed beats maintained.** You want the automation, not the infrastructure roadmap that comes with hosting it.
- **The whole team builds.** Plain-English and voice building for non-engineers, with review-before-deploy instead of a flow editor to learn.
- **Costs need owners.** Per-run pricing, workspace analytics, and a published credit rate ($25 buys 4,000).
- **Security review is coming.** Credential firewalling per operation, model-context isolation, audit logs, approval gates, and managed on-prem when required.
- **Model leverage matters.** Two dozen managed models from nine labs, effort control, auto-routing, and your own subscriptions connected in General Input.
## FAQs
### Is General Input open source?
No. If source access is a hard requirement, Activepieces is the better fit. General Input addresses the underlying concerns differently: on-prem deployment for data control, bring-your-own keys and subscriptions for model control, and exportable workflows and audit logs for exit safety.
### Can I migrate Activepieces flows to General Input?
Yes, by description: state what the flow does in the builder chat and review the result. Most flows rebuild in minutes, and gain per-run costs, approval gates, and firewalled credentials on arrival.
### Does General Input support MCP?
No. Activepieces is the stronger choice for MCP-centric architectures. General Input's integration surface is its first-party catalog: 850+ apps and 80,000+ documented operations maintained in-house.
## Better together
Some teams run both deliberately: Activepieces as the open-source, self-hosted engine for developer-owned and air-gapped flows, and General Input as the managed AI-native layer where the rest of the company builds with governance attached. A webhook bridges them in minutes.
---
# General Input vs. Polsia
> Polsia hands your company to nine autonomous agents. General Input gives you 24/7 autonomy with the approvals, audit logs, and permissions a business needs.
Polsia pushes autonomy about as far as it currently goes. Give it a goal and its nine specialized agents divide up your company: an orchestrator playing CEO and writing the daily plan, agents for social posts, outbound email, customer support, ad spend on Google and Meta, Stripe-synced finance, business planning, competitor research, and code generation that opens pull requests while you sleep. The agents share persistent memory, reach live systems over MCP, and email you a summary in the morning. Investors bought the vision to the tune of $30 million at a $250 million valuation, and the solo-founder success stories are real marketing gold.
The pitch is also the problem. An agent fleet that spends ad budget, emails prospects, ships code, and answers customers, with nobody signing off before any of it, is an arrangement that works right up until it doesn't. There is no approval step in front of the irreversible action, no audit trail an outsider can verify beyond the agents' own memory, no roles for a second human, and no way to run any of it on infrastructure you control. For a solo founder testing throwaway ideas, that trade can be rational. The moment real customers, real budgets, or a finance team enter the picture, "trust the agents" stops being a governance model.
General Input's position: you can have the 24/7 autonomy without giving up the controls. Agents run around the clock here too, but sensitive steps pause for human approval, every run is logged with inputs, outputs, timing, and exact cost, credentials are firewalled to only the operations each workflow needs, and the whole platform can deploy on-prem.
## What Polsia does well
- **Ambition made concrete.** Nine coordinating agents covering plan, build, market, support, and books is the most complete "AI runs the company" implementation shipping today.
- **Genuinely hands-off.** It really does work overnight and show you a summary in the morning, and for idea-stage experiments that speed is intoxicating.
- **Persistent shared memory.** Agents keep context across tasks and days, so the fleet's output compounds instead of resetting.
- **A vision worth taking seriously.** Autonomy at this level is where the industry is heading; Polsia is simply early, and unapologetic about it.
## How they compare
## Where they differ
### The approval gate is the product
The difference between "an agent emailed 400 prospects overnight" as a win or a crisis is whether anyone approved the list and the message. General Input lets you draw that line per workflow: enrichment runs free, anything customer-facing or money-moving pauses at an approval gate until a person clicks yes, with the full context in front of them. Polsia's model is trust-then-summary: you find out what the fleet did when it tells you. Autonomy you can scope is autonomy you can actually deploy.
### Evidence vs. memory
When finance asks what the automation spent, or a customer asks why they were emailed, General Input answers with records: every run logged with inputs, outputs, timing, tool calls, and exact cost, every credential use recorded, all exportable. Polsia's agents remember what they did, but an agent's memory is not an audit trail, and "the orchestrator's summary said so" does not survive contact with an auditor.
### Blast radius engineering
General Input assumes any model can be confused or manipulated, and engineers accordingly: credentials never enter model context, and the Agent Firewall scopes each credential to the specific operations its workflow needs, read-only, write-only, or a single endpoint, enforced at the platform layer. A prompt-injected agent cannot call what it was never given. Polsia's agents hold broad access to your ad accounts, inbox, Stripe, and repos, because broad access is what "runs your company" requires.
### One account vs. an organization
Polsia is built around one operator and their fleet. General Input is built for the org chart: workspace roles decide who can view, run, edit, or share each workflow and credential, people collaborate on the same automations, and analytics show who runs what and what it costs. Businesses have more than one human; the platform should too.
## When to use Polsia
Use Polsia when you are one person, the venture is an experiment, and speed matters more than control: idea-stage products, spin-up-and-see side projects, and ventures where the worst-case outcome is a lesson rather than a liability. At that job, nothing else is quite as hands-off.
## When to use General Input
- **Real customers or real money are in the loop.** Approval gates before the irreversible steps, and a record of everything after them.
- **More than one person is accountable.** Roles, shared workflows and credentials, and an audit trail an outsider can verify.
- **Autonomy needs a perimeter.** Credential firewalling and reviewable workflows keep the blast radius of a bad run bounded.
- **Costs must be attributable.** Exact per-run pricing at a published credit rate ($25 buys 4,000), not a monthly surprise.
- **Infrastructure is a constraint.** On-prem deployment, your own model keys, or your own subscriptions connected in General Input.
## FAQs
### Can General Input run something as autonomous as Polsia?
Yes. Agents run 24/7 on schedules, webhooks, and events, chain workflows and skills, and act across 850+ integrations. The difference is that you choose where autonomy ends: which steps run free and which wait for a human.
### Isn't approval friction the whole thing Polsia eliminates?
Approvals are per-step, not per-run: a workflow can do fifty autonomous things and pause once, before the email sends or the budget moves. Most teams find that one pause is the difference between piloting AI and prohibiting it.
### We started on Polsia. Can we move?
Yes. Describe each agent's job in General Input's builder and review the workflows it produces. The rebuilt versions gain approval gates, per-run costs, and firewalled credentials, which is usually what prompted the move.
## The bottom line
Polsia is a glimpse of full autonomy, and a fine playground for ventures where nobody gets hurt. General Input is where autonomous operations meet accountability: the same always-on agents, wrapped in the approvals, audit trail, permissions, and deployment control that let a real business say yes.
---
# General Input vs. Hermes Agent
> Nous Research's Hermes Agent rewrites itself on your own server. General Input offers governed autonomy, with approvals, audit logs, and reviewable skills.
Hermes Agent, from Nous Research, is one of the most interesting open-source projects of 2026. Launched in February under an MIT license and already past 200,000 GitHub stars, it is a self-improving agent that runs as a daemon on your own hardware, anywhere from a $5 VPS to a GPU cluster, installs with a single curl command, and reaches you through 16+ messaging platforms. Its signature is the learning loop: Hermes turns completed tasks into reusable skills, refines them during use, and carries memory across sessions, so the agent you run in December is meaningfully better than the one you installed in June. It runs on any model provider, including local models through Ollama, and shipped a desktop app this summer.
For a developer who wants an agent that grows with them, all of that is the appeal. For an organization, the same design reads differently. An agent that rewrites its own skills will not behave the same way twice, and nothing records why it made a given decision. It is built for one person on one server: no roles, no approval steps, no separation between the agent and whatever secrets live in its environment. Self-hosting keeps your data yours; the guardrails, all of them, are yours to build too.
General Input agrees with Hermes about more than you might expect: model freedom (including local models), skills as the unit of reuse, and data staying under your control. The difference is who is in charge of change. In General Input, skills are authored, versioned, and shared deliberately; sensitive actions wait for human sign-off; and every run leaves an audit trail a third party can read.
## What Hermes Agent does well
- **The learning loop.** Turning finished tasks into skills and improving them with use is a genuinely novel capability, and watching it compound is the project's magic.
- **Radical self-hosting.** MIT-licensed, one-command install, happy on a $5 VPS. Your data, your server, your rules, at a price of zero.
- **Model freedom.** Anthropic, OpenAI, Google, DeepSeek, or fully local via Ollama. No lock-in at the model layer.
- **Meet-you-anywhere interface.** A daemon with cron scheduling you can talk to from 16+ chat platforms is a great personal-agent form factor.
## How they compare
## Where they differ
### Who authors the skills
Both platforms bet on skills as the unit of reusable capability. Hermes writes its own: the agent distills experience into skills and edits them as it learns, which is the point, and also the audit problem, because the system's behavior is a moving target no one reviewed. General Input's skills are the opposite kind of asset: authored deliberately (by people, or by AI with a person approving), versioned, shared across workflows and people under permissions, and identical on every run until someone changes them on purpose. One is evolution; the other is engineering. Organizations need to know which one is running their process.
### The blast radius question
Hermes runs with the keys of the environment it lives in: whatever tokens and files sit on that server are, in practice, available to the agent. On your personal VPS, that is a reasonable trade. General Input is engineered for the opposite assumption: workflow code runs in isolated sandboxes, credentials are encrypted and never enter model context, and the Agent Firewall scopes each credential to only the operations its workflow needs, read-only, write-only, or a single endpoint, enforced at the platform layer where a prompt-injected model cannot reach. Every credential use is logged.
### Autonomy with a pause button
Hermes executes what you ask, and its cron jobs run unattended by design; there is no built-in step where a human approves the risky action before it happens. In General Input, approval gates are a first-class workflow step: the enrichment runs free, the customer-facing email waits for sign-off, and you choose which is which per workflow. That single design decision is what separates a personal agent from an automation platform a company can adopt.
### Sovereignty without the server admin job
Hermes' answer to data control is "run it yourself." General Input offers the same end without the ops: connect local models on your own hardware to workflows so prompts run on compute you control, or connect your existing AI subscriptions directly in the product; bring-your-own keys cover the API path; and for full data residency, the Enterprise on-prem deployment puts the whole platform on your infrastructure with the vendor carrying the operational load alongside you.
## When to use Hermes Agent
Use Hermes when you are a developer or tinkerer who wants a personal agent that compounds: your server, your models, your experiments, and a system that visibly gets better the more you use it. It is also a fascinating base for skill-learning research. For a personal sidekick you fully own, it is arguably the most interesting option available.
## When to use General Input
- **The automation is a business process.** Reviewable workflows, deterministic code steps, and behavior that does not drift between runs.
- **Someone else has to trust it.** Approval gates, exportable audit logs, and per-run cost records that satisfy finance, security, and customers.
- **A team is involved.** Roles, shared credentials, shared skills, and workflows more than one person can safely own.
- **Secrets need a perimeter.** Sandboxed execution and per-operation credential firewalling instead of environment-level access.
- **You want the reach without the wiring.** 850+ managed integrations with OAuth handled, versus building connectivity skill by skill.
## FAQs
### Does General Input support local models like Hermes does?
Yes. General Input supports local models running on your own hardware and lets you connect your existing ChatGPT, Claude, Gemini, or Grok subscriptions to your workflows. The managed catalog also includes open-weights models like DeepSeek V4, Kimi K3, and GLM.
### Are General Input's skills self-improving?
Not autonomously, by design. Skills improve when a person (or an AI drafting for a person's approval) updates them, so every change is reviewable and attributable. You get compounding capability without unreviewed drift.
### Can I use both?
Plenty of developers do: Hermes as the personal agent on a private VPS, General Input for the automations that touch company systems, customers, or money, where governance is the requirement.
## The bottom line
Hermes Agent is the most compelling personal agent in open source: self-hosted, self-improving, and free. General Input is what the same instincts look like when they grow up into infrastructure: model freedom and skills, plus the approvals, audit trail, credential firewall, and team permissions that let an organization, not just a developer, put agents to work.
---
# General Input vs. OpenClaw
> OpenClaw gives an LLM the run of your machine. General Input runs automations in sandboxes with firewalled credentials, approvals, and audit logs.
OpenClaw is a phenomenon. Peter Steinberger's open-source personal assistant rocketed past 200,000 GitHub stars, built a huge community, and earned the affection its "Molty" nickname implies. The design is radically direct: a local gateway that gives an LLM real access to your machine, reading and writing files, running scripts, controlling a browser, and reaching you through WhatsApp, Telegram, Slack, iMessage, and more. With 50+ integrations and a growing skills ecosystem, it will clear your inbox, draft your posts, review code, and even fill in job applications. It is free, it runs on your hardware, and it is genuinely fun.
It has also become the case study in why "an LLM with a shell" alarms security teams. 2026 brought CVE-2026-25253, a one-click remote code execution flaw scoring 8.8 on CVSS, plus further high-impact advisories for command injection. Credentials sit in a local JSON file the model's tooling can reach. And the project explicitly scopes prompt injection attacks out of its security fixes, which means the core attack vector against an agent that reads untrusted content with system access is, by policy, your problem. On your own laptop, that is a personal risk decision. The moment company data, customer systems, or work credentials are involved, it is the part that fails the security review.
General Input is built on the opposite premise: assume the model can be manipulated, and make that survivable. Workflow code runs in isolated, disposable sandboxes. Credentials are encrypted, never enter model context, and are scoped by the Agent Firewall to only the operations each workflow needs. Risky actions wait for human approval, and every run leaves an exportable audit trail.
## What OpenClaw does well
- **Frictionless personal power.** Your files, your apps, your browser, one chat message away. For a power user's own machine, the capability-to-effort ratio is unmatched.
- **Meets you in your messages.** WhatsApp, Telegram, Slack, iMessage, and more. No new app to open, ever.
- **Community velocity.** A massive contributor base, a thriving skills ecosystem, and rapid releases (including security patches when issues surface).
- **Free and local.** No subscription, no cloud dependency, full ownership of the stack.
## How they compare
## Where they differ
### The perimeter
OpenClaw's power comes from having none: the agent operates your actual machine with your actual permissions, so everything you can touch, it can touch. General Input inverts this. Workflow code executes in throwaway sandboxes isolated from any machine you care about; the worst a bad instruction can do is bounded by what that one workflow was granted. The 2026 CVEs made the difference concrete: a one-click RCE in a tool that has the run of your laptop is a very different event than a compromised run inside a disposable sandbox holding one firewalled credential.
### Secrets and the injection problem
Every agent that reads email, web pages, or documents ingests attacker-controllable text. OpenClaw's docs are honest that prompt injection defenses are out of scope, and its credentials live in a local file within the agent's reach, so a successful injection can act with everything the assistant has. General Input assumes injection will eventually happen and engineers for it: credentials never appear in model context, and the Agent Firewall enforces per-operation scope, this key can read the CRM but never delete, refund but never pay out, at the platform layer the model cannot negotiate with. Every credential use is logged.
### Asking first vs. apologizing after
OpenClaw executes when you ask, including actions you cannot take back: the sent email, the deleted file, the submitted form. General Input makes "pause here for a human" a first-class step. Sensitive workflows stop at an approval gate with full context, and continue only on sign-off. Combined with review-before-deploy, structured workflows you inspect before they ever run, you know what an automation will do before it does it, not after.
### One person's machine vs. an organization
OpenClaw is gloriously single-player: one human, one gateway, one set of powers. General Input is built for the org: workspace roles over every workflow and credential, sharing with collaborators, per-run cost accounting at a published credit rate ($25 buys 4,000), workspace analytics, and an audit trail an outside party can verify. And where OpenClaw's model access is whatever keys you give it, General Input's catalog spans two dozen managed models from nine labs, plus your existing AI subscriptions and local models.
## When to use OpenClaw
Use OpenClaw as a personal power tool on your own devices, with your own risk budget: inbox triage, personal projects, home automation, browser chores. Isolate it from work credentials, keep it patched, and it is one of the most fun tools in AI right now.
## When to use General Input
- **Company data or customer systems are involved.** Sandboxed execution and firewalled credentials keep a bad run from becoming an incident.
- **The security team gets a vote.** Model-context credential isolation, per-operation scoping, audit logs, and approval gates are the review-passing answers.
- **Irreversible actions need a human.** Approval gates before sends, deletes, and payments.
- **A team shares the automation.** Roles, credential sharing, and workflows that outlive any one person's laptop.
- **You need reach with accountability.** 850+ managed integrations, exact per-run costs, and an exportable record of everything.
## FAQs
### Is OpenClaw actually unsafe?
It is honest about its trade: maximum capability, user-managed risk. The 2026 CVEs were patched, but the architecture, direct machine access, reachable credentials, out-of-scope prompt injection, is the design, not a bug. Fine for a personal machine; hard to defend for work systems.
### Can General Input control a browser like OpenClaw?
No. General Input works through APIs: 850+ integrations with 80,000+ documented operations. For automations that only exist as clicking through a UI, OpenClaw (on a personal machine) is the tool; for production reliability, the API path wins.
### Can I message General Input like I message OpenClaw?
The primary interfaces are the dashboard chat (with voice dictation) and triggers, including email. It does not sit in WhatsApp or iMessage the way OpenClaw does.
## Better together
Run OpenClaw as your personal sidekick on your own hardware, and General Input as the automation layer for anything that touches work: shared systems, customer data, credentials, and money. The line between them is exactly the line your security team would draw anyway.
---
# How Newtex eliminated hours of daily export paperwork with General Input
> Newtex's AI specialists produce export documents in seconds, replace hand-coded financial reports, and free the team for revenue-generating work.
Customer story. Published: 2026-10-03. PDF: https://www.generalinput.com/customer-stories/newtex/Newtex-Customer-Story.pdf
- **99% faster exports** from 12 minutes per document to seconds
- **40+ hours/month saved** on export paperwork the team used to retype by hand
> "By rapidly automating repetitive (but critical) tasks in Sales and Operations with General Input, I expect my team to redeploy 30–40% of their time to revenue-generating activity."
>
> **Jerry Joliet**, CEO, Newtex Industries
- **Company name:** Newtex Industries
- **Industry:** Manufacturing
- **Company size:** Mid-market
- **Pain point:** Manual export paperwork and hand-coded reports
- **General Input products used:** AI Specialists, Apps, Workflows, Integrations
- **Connected systems:** ERP database, BI reporting tool, Google Drive, Gmail, payroll system, CRM
Background
## A mid-sized manufacturer shipping to customers around the world
Newtex Industries manufactures high-temperature textiles and protective products in Victor, New York, and ships them to customers around the world. Like a lot of mid-sized manufacturers, it runs on an ERP that does the core job well and makes everything around it hard: forms, reports and the paperwork that moves between them.
In January, CEO Jerry Joliet made AI proficiency a major objective for Newtex’s senior staff. The team now runs AI specialists in General Input across sales, operations and finance. Export documents that used to be retyped by hand come together in seconds, and financial reports that once needed custom code are a conversation away.
The problem
## Fifteen-year-old spreadsheets and reports that took code to change
Every international order Newtex ships needs three documents before it leaves the building: a pro forma invoice so the customer can arrange shipping and customs, a commercial invoice and packing list for customs and the carrier, and a shipper’s letter of instruction for the freight forwarder. 90% of the information is the same across all three.
The ERP could print an order confirmation, but changing its forms meant working in old code. “Coding a form in any ERP is a nightmare,” says Jerry. “It’s old code. It’s not drag and drop.” So fifteen years ago Newtex built its own forms in Excel. They looked professional, but nothing in them was automated, and every export order was retyped by hand.
Each document took 10 to 12 minutes. At four export orders a day and three documents per order, that added up to two hours of skilled time, every day.
Reporting had the same problem. Leadership reviews the financials a couple of times a month to track performance, and the reports lived in a business intelligence tool that was powerful but hard to set up. “It was hard,” says Jerry. “I had to actually code them in some cases.”
Outside help hadn’t fixed it either. Over the last decade, Newtex hired three consultants who were supposed to be experts to optimize its ERP planning process. None of them delivered any value, and Newtex’s own people often ended up teaching the consultants how the business worked.
Meanwhile the team was stretched. Newtex’s people are process-driven and focused on productivity, and they embrace anything that makes the work faster and more efficient. What they lacked was a tool that could keep up.
The solution
## AI specialists that build export documents in seconds and financial reports on request
### Export documents, built in under an hour
Exports came first. Jerry wanted to see how fast the work could be automated, so he built the beta himself. He created an Export Coordinator specialist and gave it one sentence of instructions: manage export paperwork, look up and store HTS codes, find tariffs, and create the pro forma invoice, commercial invoice and shipper’s letter of instruction. He pointed it at the existing forms in Google Drive and a recent export order, and asked it to build the documents from that order. “I literally did this in less than an hour,” he says. Then he handed it to Lindsey, Newtex’s VP of Sales, who tested and tweaked it and is rolling it out to her team.
Now someone on the team enters a customer order number and the app builds all three documents. Items, prices, parties and terms come straight from the ERP and stay locked. Shared details like incoterms, weights and HTS codes are entered once and flow into every document, and a checks panel lists exactly what a person still has to add or confirm before anything is downloaded.
The documents look like Newtex, too. “I don’t even think we gave it design instructions,” says Jerry. “We just said, these are our colors, this is our logo, make it look good.”
### Financial statements, built by conversation
Finance came next. The financial statements app in the Accounting Manager specialist was built “by giving it some structured data and talking to it,” says Jerry. It read the spreadsheets Newtex had already built and mapped the general ledger to the statement lines finance uses.
The income statement compares the current period to last year with percent-of-sales and change columns. Cost of goods sold breaks down into material, labor, overhead and variance, EBITDA is built up with Newtex’s standard adjustments, and the app writes notes on the largest items. When finance wants something new, it asks in the same conversation.
### A specialist for every department
Exports and finance were only the start. Newtex now has a specialist for every department, each set up like a new hire for one part of the business, with its own files, apps, connections and conversation history. Each app is shared with the people who use it every day.
The results
## Hours back every day, redeployed to revenue
Export documents that took 10 to 12 minutes each now come together in seconds, and a person reviews only what the checks flag. “Is it ready to go? No,” says Jerry, “but it’s 99% of the way there.” On the finance side, leadership gets an income statement, EBITDA bridge and notes whenever they need them. “Right now the app is actually much easier to use than [our BI tool],” he says, and Newtex is now building toward replacing that tool entirely.
The same idea runs through every specialist: the software does the checking, and people look only at what’s wrong. “Instead of doing all the work that we used to do, we’re only going to look at exceptions,” says Jerry.
The whole team uses General Input. Lindsey, the VP of Sales, is using it to review what each customer does and buys, find companies in the same applications that Newtex doesn’t sell to yet, and build campaigns that bring in inbound leads. “She’s learned AI, and her skills are expanding daily,” says Jerry. “I have employees that are having fun with it, because it’s just mind-boggling how quickly you can do it.”
Next, Newtex is moving its ERP and payroll data into a central data warehouse and building out invoice matching and production planning. “I expect every business process that runs on spreadsheets, Drive or Google Forms to be rebuilt in General Input,” says Jerry.
## FAQs
### What did Newtex build with General Input?
A specialist for every department. The Export Coordinator turns an ERP sales order into a pro forma invoice, a commercial invoice and packing list, and a shipper's letter of instruction. The Accounting Manager produces live financial statements and an AP aging tracker. Others check supplier invoices, report payroll hours, trace bills of materials and research new leads.
### How long did it take to set up?
Jerry built the first version of the export documents app in under an hour, from a one-sentence description, an example order and Newtex's existing forms. Newtex's VP of Sales then tested and refined it for her team.
### Does a person still review the export documents?
Yes. Items, prices and parties come straight from the ERP order and stay locked. A document checks panel lists what a person still has to add, such as package counts and weights, and what to confirm, such as HTS codes and export classification, before anything is downloaded.
### Which systems does it connect to?
Newtex's specialists work with its ERP database, its BI reporting tool, Google Drive, Gmail, its payroll system and its CRM.
---
# How Primis became an AI-native recruitment firm with General Input
> Primis gives every consultant a daily lead list and morning briefing, and cut monthly commission from two days to 20 minutes with AI specialists.
Customer story. Published: 2026-10-03. PDF: https://www.generalinput.com/customer-stories/primis/Primis-Customer-Story.pdf
- **300+ hours/month saved** on lead research and meeting prep for consultants
- **98% faster commission** from two days of spreadsheets to a 20-minute run
> "Every consultant now starts the day with a briefing and a fresh list of leads, and commission that took two days takes twenty minutes. We set out to be the best AI-enabled recruitment business in the world, and this is how we’re getting there."
>
> **Ben Broughton**, Founder & CEO, Primis
- **Company name:** Primis
- **Industry:** Recruitment and staffing
- **Company size:** 50 employees
- **Offices:** London, Manchester, New York, Dubai
- **Pain point:** Manual research, reporting and back-office admin
- **General Input products used:** AI Specialists, Apps, Workflows, Integrations
- **Connected systems:** Recruitment CRM, Microsoft 365, e-signature, HR system, timesheet system, phone system
Background
## A fast-growing tech recruitment firm that wanted to lead on AI
Primis is a technology recruitment firm headquartered in London, with offices in Manchester, New York and Dubai. Founded in 2021, it places permanent and contract talent across engineering, data, AI and go-to-market roles, and was named Recruitment Company of the Year for 2025 by the industry body APSCo.
Founder and CEO Ben Broughton made AI the firm’s top priority. “If you’re not AI-enabled, you can get left behind,” he says. He met with 12 companies before choosing General Input, and within weeks the whole firm was working in its own branded AI workspace.
The problem
## Great consultants, slowed down by research, spreadsheets and copy-paste
Recruitment runs on knowing which companies are hiring before anyone else does. At Primis, consultants mostly found out by accident. “We’re almost stumbling across, or hearing from candidates about, these companies that are hiring,” says Simon Gregory, who leads the UK go-to-market desk. “By then you’re almost too late.” Researching each company and drafting outreach was manual, and it took time.
General-purpose chatbots couldn’t do the job at scale. When Ben asked one to research every company Primis had placed with and find their live roles, it came back with: “I can’t do this. I’ll give you the top 10.” And the firm’s own data was hard to reach. Its CRM held 40,000 companies and years of placement history, but “we’re finding it hard to get the real-time data out,” Ben says.
The back office ran on spreadsheets. Every month, the finance team spent two days calculating commission by hand, moving between three or four spreadsheets and the CRM to check splits. “It’s just a lot,” says Safeera Pillay, who runs the numbers. Contractor contracts were copied from a form into documents by hand, and onboarding a new hire meant working through a long manual checklist.
The solution
## AI specialists for every desk, from business development to finance
### A fresh lead list for every consultant, every morning
Primis started with the work that grows the business. A Lead Generation Executive specialist builds a personal lead list for every consultant each weekday morning. It looks for companies with real hiring signals in that consultant’s market, like headcount growth, senior hires, open roles and new funding, and checks each one against the CRM and against colleagues so no two consultants get the same lead.
Consultants work their list in the Lead List app: they rate leads so the specialist learns what they like, find the hiring manager, add the company to the CRM in one click and open a drafted outreach email. Every consultant on the desk starts the day with a fresh list, 175 new leads across the firm every weekday. “Each day I get a fresh list, and 60 to 70% of them are genuinely interesting,” says Simon.
### Commission in 20 minutes, not two days
In finance, a Commission Payout Specialist took over the monthly commission run. The team uploads the month’s placement and contractor reports and sets exchange rates, and the Commission Engine produces the full workbook, with splits, management schemes and a tab for every consultant. Anything that doesn’t match the CRM is flagged for a person to check, and nothing is paid until someone signs it off.
The first run on real July figures took 10 minutes and matched the team’s own numbers to within a dollar or two. “Was that just 10 minutes?” Safeera asked. “My job’s going to be so much easier next month.” Commission now takes 20 minutes and a final check, down from two days.
### A specialist for every part of the firm
Primis now has a specialist for every part of the firm: a call analyst, a revenue analyst, interview reminders, an inbox that files emailed CVs straight into the CRM, and a report that flags every live vacancy with no recent activity.
The morning briefing runs for the whole team, each copy tuned to its owner’s calendar and start time.
The results
## A firm that runs on AI, from the CEO to the newest consultant
It runs across the whole business. Consultants start every day with a researched briefing and a fresh lead list, finance closes commission in twenty minutes, and operations hands contracts and onboarding to specialists that never skip a step. Between lead research and meeting prep alone, that is more than 300 hours of consultant time back every month.
The work runs on its own. 97.6% of automated runs succeed, and 98.6% start on a schedule or trigger, with nobody pressing a button.
And the time is going where it counts. “We’re looking at saving time and making people more efficient, but also delivering more business to our front door,” Ben says. Next, Primis is automating board packs, monthly consultant reviews and follow-up sequences.
## FAQs
### What did Primis build with General Input?
A team of AI specialists. A Lead Generation Executive sends every consultant a personal lead list each morning, a Commission Payout Specialist produces the monthly commission workbook, and others handle contractor contracts, employee onboarding, CV formatting, candidate matching, call summaries and live reporting from the recruitment CRM.
### How long does commission take now?
20 minutes and a final check, down from two days of spreadsheet work each month. The first run on real figures matched the finance team's own numbers to within a dollar or two, and nothing is paid until a person signs it off.
### Who at Primis uses it?
The whole firm. Consultants use it for leads, morning briefings and CVs, finance for commission, and operations for contracts and onboarding.
### Which systems does it connect to?
Primis's specialists work with its recruitment CRM, Microsoft 365 email and calendar, its e-signature tool, HR system, timesheet system and phone system. Primis runs General Input as its own branded workspace.
---
# How Nove turns every bulk inquiry into a quote in 2.5 minutes with General Input
> Nove's AI specialists quote bulk orders in minutes, follow up until buyers reply and draft every support email, so the founder can focus on the product.
Customer story. Published: 2026-10-03. PDF: https://www.generalinput.com/customer-stories/nove/Nove-Customer-Story.pdf
- **40 hours/month saved** for a founder who spent half his day on email
- **2.5-minute quotes** on every bulk inquiry, nights and weekends included
- **5-star reviews** for order fixes handled with 0 manual intervention
> "Half my day used to go to customer emails. Now every bulk inquiry gets a quote within minutes and support replies are drafted before I open my inbox. It just works."
>
> **Brendan Mudd**, Founder, Nove
- **Company name:** Nove
- **Industry:** Consumer products, direct and wholesale
- **Company size:** Founder-led, with part-time freelancers
- **Location:** San Francisco, California
- **Website:** [www.novedevice.com](https://www.novedevice.com)
- **Pain point:** Quotes, follow-ups and customer email handled by hand
- **General Input products used:** AI Specialists, Workflows, Skills, Apps
- **Connected systems:** Online store, accounting system, CRM, email, warehouse, file storage
Background
## A founder-run company selling to institutions across the U.S.
Nove is a San Francisco company that designs a discreet keychain carry case for single-dose emergency nasal sprays. It sells direct to consumers and in bulk to health departments, universities and community programs, and its website, [www.novedevice.com](https://www.novedevice.com), cites more than 100 institutional clients across the U.S.
Founder Brendan Mudd, an industrial designer, runs the business himself with a few part-time freelancers. Every bulk order starts with an inquiry, a quote and a run of follow-up emails, and for a long time Brendan handled most of that in the gaps around a full-time job.
The problem
## A wholesale pipeline held together by brittle automations and a busy founder
Nove’s quote process ran on a chain of four separate no-code automations that a part-time contractor maintained. They worked, but they were rigid. Follow-ups went out on a fixed timer with no context, and Brendan often had to send a second note apologizing for the first. “The message sounds kind of out of touch and out of context,” he says, “and then I have to follow up and say, sorry about our automated message.” Changing anything meant rebuilding it from the start, and waiting until the contractor had time.
After the quote, everything was manual: tax-exemption forms, revised quotes, invoices and vendor paperwork, all by email. Customer questions about orders, addresses and pricing filled the rest of the day. “Honestly, half my day now is just responding to customer emails,” Brendan said. “I’m managing orders and placing orders, and I just don’t need to be doing that.”
He couldn’t keep running it that way. “I can’t just manage it anymore in my lunch breaks and a couple of hours before and after work,” he said. “It’s a bit of a bottleneck.”
The solution
## A wholesale assistant that quotes, follows up and knows when to stop
### Quotes in minutes, follow-ups that know when to stop
In August, Nove replaced its old no-code chain in a single afternoon. Now a Wholesale Assistant watches the bulk-order inbox. When an inquiry arrives, it creates an estimate in the accounting system and emails it to the buyer, a median of 2.5 minutes later. 43% of those quotes went out in the evening or at the weekend.
Then it keeps the deal moving. A follow-up sequence checks the CRM every weekday and writes to buyers who haven’t replied after two, five and twenty days. 26% of those follow-ups get a reply, and it marks each of those leads as in progress so they can be picked up right away. And it knows when to stop: it held back follow-ups on estimates marked tax-exempt with no certificate on file, and held a quote when the shipping address was too incomplete to calculate tax.
### Support replies drafted before anyone opens the inbox
Customer email came next. A Customer Service specialist reads every contact-form message and drafts a reply using Nove’s FAQ and live order data, so answers about order status, address changes and bulk pricing are ready a median of 2 minutes after the email arrives. Every reply is reviewed before it goes out. “It took a couple of minutes, and it just worked,” he says.
Because every workflow is described in plain English, changing a process takes a sentence, not a rebuild. When Nove redesigned its quote form, the workflow adjusted on its own. “I completely changed the form, and it just picked it up and adapted everything automatically,” he says.
### One founder, a team of specialists
The work happens in chat with the specialists, and anything that repeats becomes an automation.
### Beyond sales: support, operations and finance
The same approach now covers most of the business. Each specialist handles a job the founder used to do by hand, and checks in before anything that needs judgment.
The results
## A wholesale business that keeps selling while the founder works on the product
Nove’s specialists now turn out quotes, follow-ups, support drafts, digests and audits every week, and every run starts on its own, from an email arriving or a schedule. That is about 40 hours a month back for a founder who used to spend half his day on email.
The work is reliable. 96.8% of Nove’s automated runs succeed, and the only deliberate stops were the agent holding a quote or follow-up for a person to check. Customers notice the difference: one left a 5-star review about how quickly Nove fixed a problem with their order, a fix handled end to end by a General Input specialist.
Next, Nove is adding social messaging for creator outreach and order questions, and building a portal where institutional buyers can reorder on their own.
## FAQs
### What did Nove build with General Input?
A team of AI specialists and automations. A Wholesale Assistant turns bulk inquiries into emailed estimates and follows up until buyers reply, a Customer Service specialist drafts replies to every contact-form email, and others handle a daily digest, inbound shipments, product costs and an accounts check.
### How fast are quotes sent?
A median of 2.5 minutes after a bulk inquiry arrives, evenings and weekends included. 43% of quotes go out outside business hours.
### Does a person still review the work?
Yes. Support replies are drafts that Brendan reviews and sends. The wholesale assistant holds a quote or follow-up for Brendan whenever something needs judgment, such as a missing tax certificate or an incomplete address.
### Which systems does it connect to?
Nove's specialists work with its online store, accounting system, CRM, email, warehouse and file storage.
---
# Build Your AI Workforce with General Input Specialists
> General Input Specialists are live on the web and iPhone. Describe any task, and a specialist learns how you do it and takes it off your plate.
Published: 2026-10-13. Last updated: 2026-10-13.
Today, General Input Specialists are open to everyone, on the web and on iPhone.
Our goal with General Input is simple: automate the parts of your job you hate doing, so you can spend more time on the work you love. Specialists are how we get there. Each one is an AI agent that learns one part of your work, does it the way you would, and takes more of it off your plate every time you delegate.
## Describe the task, and a specialist takes it from there
Start by describing any task you'd normally do on your computer, in your own words. Pull last month's churned accounts and draft a win-back email for each one. Turn this signed SOW into a contract from our template. Check the supplier portal every morning and enter new purchase orders in our ERP.
If you already have a specialist trained for that work, the task goes straight to it. If you don't, General Input creates a new specialist to become the expert, and it asks what it needs to know before it starts.
Then it keeps what it learned. A specialist writes down how you like the work done, as skills and memory in plain language, so the next request starts from your process instead of from scratch. The more you delegate, the more it can take on without being told twice.
## What specialists can do
Specialists do the work, not just describe it.
- **Build dashboards and apps.** An operating report your team opens every morning, or a tool built around a process only your company has.
- **Create documents.** Contracts, reports, and proposals, formatted and ready to send.
- **Browse the web.** A specialist works in a real browser, so it can use the sites and legacy systems that have no API, and you can watch it work live.
- **Connect to the tools your team already uses.** Gmail, Slack, HubSpot, Stripe, DocuSign, and the rest of your stack.
Work can also run on a schedule or when something happens, so it keeps going between conversations.
## Built for your whole team
Any specialist can be shared with your entire team. Everyone gets the same expert with the same skills and memory, so the way your best person handles a job becomes the way your company handles it, instead of staying trapped on one person's laptop.
Your conversations stay private. What you share is what the specialist knows and does: its skills, memory, files, and the work it runs.
## No limits on your workforce
Unlike other solutions, General Input doesn't cap how far you take it.
- **Every model.** Use any major model, including the ChatGPT or Claude subscription you already pay for.
- **As many specialists as you need.** One job per specialist works best, for the same reason [a dozen small agents beat one giant one](/blog/small-agents-not-mega-skills), so create one for every job you have.
- **No limit on concurrent executions.** Workflows run in parallel, so one long job never holds up the rest.
## Your specialists, on your phone
The General Input iPhone app launches today. Hand a specialist a task from anywhere, get a notification when it needs a decision, approve its work in a tap, and open the apps it builds. Everything stays in sync with the web, so a conversation you start at your desk picks up on your phone.
Get it at [generalinput.com/mobile](/mobile).
## Start with a free month
Every new workspace comes with a free month of credit to start building your own digital workforce. Start with the task you hate most. If you'd rather begin from a working example, [browse specialist templates](/templates/specialists), each built around a single job.
For the thinking behind specialists and what changed for existing workspaces, read [our first post on specialists](/blog/introducing-ai-specialists). And if you want help setting up your first one, [book a call with Jack](https://cal.com/jackjoliet/concierge-support).
## FAQs
### What is a General Input specialist?
A specialist is an AI agent that learns one part of your work, such as reporting, contract drafting, or lead follow-up, and does it for you. It builds and runs the dashboards, documents, and workflows that work needs, and keeps what it learns about your process as skills and memory your team can read.
### Do I have to set up a specialist before I give it work?
No. Describe the task in your own words. If a specialist already handles that kind of work, the task goes straight to it. If not, General Input creates a new specialist to become the expert.
### Which AI models do specialists use?
General Input works with all major models, including the ChatGPT or Claude subscription you already pay for.
### Can my team use the same specialists?
Yes. Any specialist can be shared with your whole team, so everyone works from the same skills and memory. Your conversations with a specialist stay private to you.
### Is there a mobile app?
Yes. The iPhone app lets you give specialists work, answer their questions, approve what they do, and open the apps they build. Get it at generalinput.com/mobile.
### How much does it cost to get started?
Every new workspace starts with a free month of credit, and no credit card is required. There's no limit on how many specialists you create or how many workflows run at once.
---
# Introducing General Input Specialists
> General Input now runs on AI specialists, one for each job in your business. Here's why we stopped asking you to build automations and let you delegate.
Published: 2026-09-27. Last updated: 2026-10-13.
Update, October 13: specialists are now open to everyone, on the web and on
iPhone. [Read the launch post](/blog/build-your-ai-workforce).
General Input now runs on AI specialists. Each specialist owns one job in your business, such as weekly reporting, lead follow-up, or customer win-back. You tell it what you want done in your own words, and it builds and runs the skills, workflows, and apps behind that job. When the job changes, you tell the specialist, the same way you would tell the person who owns it.
If you already use General Input, everything you built is still here. The difference is that every piece of it now has an owner you can talk to.
## We were asking you to build software
People come to General Input to get work off their plate. Until now, the first thing we asked in return was a set of builder's questions. Should this be a workflow or an app? What triggers it? What should it be called, and where should it live?
Anyone who has handed a job to a capable colleague knows delegation doesn't work like that. You describe the result you want, answer a few questions, and let them figure out the rest.
Specialists work that way. Tell one, "Every Monday, send me revenue by plan and flag anything that dropped more than 10%." It asks where your revenue data lives and who should get the report. Then it builds what the job needs: a skill that records how the report is put together, a workflow that runs it every Monday morning, and an app where your team reads the results. You describe the job once, and the specialist works out which pieces it takes.
## Automation needed an owner
The systems that worked best on General Input were never a single workflow or app. They were a few of each, working toward one goal, and they stayed useful because someone looked after them. That person noticed when a report came out wrong, remembered how the pieces fit together, and updated them when the process changed. Usually it was you.
A month ago we launched [projects](/blog/introducing-projects) to give those systems a home. Projects kept things organized, but a folder can't notice a failed run or ask what you meant. They also left an awkward question: who do you talk to? You could ask a general assistant, the assistant on each workflow and app, or a chat inside the project, and each one saw only part of the picture.
A specialist answers both. It owns the job, so there is one place to ask for a change, and it's the one that comes to you when something needs a decision. Its conversations, memory, files, and work all live on one page.
## Why AI specialists build systems you can see
There's a version of AI delegation where an agent does the job from scratch every time, out of sight. It demos well, and it's hard to trust with anything that matters, because nobody can check how the work was done.
A specialist turns what you teach it into things your team can inspect. Skills write down how the job is done. Workflows run on schedules and triggers the same way every time, and every run is recorded. Apps put the results where everyone can see them. And the specialist's memory, which holds its rules and the decisions behind them, is plain language that everyone you share it with can read.
That's also what separates a specialist from a chatbot. A chatbot works while you prompt it and forgets the job when the conversation ends. A specialist's work keeps running between conversations, and what it learned about your process stays written down, so the process no longer lives in one person's head.
## Why one job each
We call them specialists because each one is meant to do one job well. We've written about why [a dozen small agents beat one giant one](/blog/small-agents-not-mega-skills): small ones are easier to understand, to hand to a colleague, and to fix when the work changes. So when a new job comes up, it gets its own specialist.
Every member starts with one, their Chief of Staff. It holds anything that doesn't have an owner yet and can set up a new specialist when a job deserves one. If you'd rather start from a working example, [browse specialist templates](/templates/specialists), each built around a single job.
## What changes if you already use General Input
Nothing you built was deleted or turned off. Each project is now a specialist, with the same workflows, apps, skills, and files, and everything outside a project is with your Chief of Staff.
The next time you sign in, a short tour walks you through the change. If you have work to sort, it offers to split it into specialists, one per job, and shows you the proposed split before anything moves.
If you have questions or want help setting up your first specialist, [book a call with Jack](https://cal.com/jackjoliet/concierge-support).
## FAQs
### What is an AI specialist?
An AI specialist is an AI agent that owns one job in your business, such as weekly reporting or lead follow-up. In General Input, it builds and runs the skills, workflows, and apps that job needs, keeps a written memory of how the work is done, and is who you talk to when the work should change.
### How is an AI specialist different from a chatbot?
A chatbot works only while you prompt it and forgets the job when the conversation ends. An AI specialist keeps its work running on schedules and triggers, writes down how its job is done, and comes to you when something needs a decision.
### How is an AI specialist different from workflow automation?
With workflow automation, you design every step and trigger yourself. With a specialist, you describe the outcome, it asks what it needs to know, and it builds and maintains the workflows, apps, and skills behind it.
### What happened to projects in General Input?
Every project became a specialist, with the same workflows, apps, skills, and files, so nothing needs rebuilding. Work that wasn't in a project is with your Chief of Staff, and a short tour offers to sort it into specialists, showing you the split before anything moves.
### How much work should one AI specialist handle?
One job, the kind of work one person would own at a desk: a channel, a workstream, a recurring report. When a new job comes up, give it a new specialist.
### Can my team see my conversations with a specialist?
No. Your conversations are private to you, even when the specialist is shared. What it knows lives in its memory and files, which everyone you share it with can read.
---
# Introducing Projects: Skills, Workflows, and Apps That Live Together
> Projects are live in General Input. Group the skills, workflows, apps, and files that serve one goal into one system, with its own memory and sharing.
Published: 2026-08-31. Last updated: 2026-08-31.
Skills, workflows, and apps are the three things you build in General Input. Until now they all lived in one flat pile.
That is fine for the first few. It stops being fine once you have a goal that takes several of them: a skill holding the procedure, a workflow running it on a schedule, an app where someone reads the output, and the reference files behind all three. Four things that only make sense together, scattered across four lists.
## Projects Are Now Live
A project is a container for one goal. It holds workflows, apps, and skills, plus its own files and its own chats, and it gives that whole system a single home you can open, extend, and hand to someone.
Create one from **Projects** in the sidebar, give it a name and an icon, and you have somewhere for the work to accumulate.
## Describe the System, Not the Pieces
The fastest way to fill a project is to skip the pieces entirely and describe what you want the system to do.
Open a chat from inside the project and it gets pinned to the input, so everything created in that conversation lands in the project instead of loose in your workspace. Then say what you want in plain English.
In the video above, the ask is one paragraph: read the previous day's news at 3am, generate five blog post ideas from it, learn our writing tone from the posts already on our blog, and give us somewhere to work the ideas.
What comes back is three pieces, wired together: a skill holding the tone, a workflow on a 3am schedule, and an app to triage the output. That took about fifteen minutes, and it is a first prototype rather than a finished thing.
A useful shape for that first paragraph: what you want produced, where the raw
material comes from, when it should run, and what it should imitate.
Extending it later is another sentence typed into the same project. Adding search data or saving images from each article is not a new build, because there is somewhere for the change to land.
## Every Project Has Its Own Memory
A project keeps a memory separate from your workspace memory and your personal memory. It builds that from the project's files, the workflows pinned to it, and how those runs go, so the system gets better at its job without you maintaining a prompt by hand.
Conversations you have in a project stay private to you, even when the project
itself is shared. What the project's specialist knows is written in its memory
and its files, in the open.
## Share the System, Not the Pieces
Sharing a project shares what is in it. Give someone in your workspace access and the workflows, apps, and skills pinned to it come along, which is usually what you want when you are handing over a system rather than a file.
Access set directly on a resource still wins. Somebody can be an editor on the project and a viewer on one particular workflow inside it.
## Deleting a Project
Deleting a project asks what should happen to the work inside it. **Delete them too** removes everything you own along with the project. **Keep them in my workspace** moves it all back out to the top level, where it keeps working, and only the project goes away.
Anything someone else owns is never deleted. It moves to your workspace instead.
## A Few Projects Worth Building
- **Content engine.** A skill holding your writing tone, a workflow that reads the news each morning and drafts ideas, and an app to approve or kill them.
- **Client onboarding.** The steps written down as a skill, a workflow that creates the accounts and folders, and an app showing where every new client is in the process.
- **Engineering health.** Your data model as a skill, bug-triage workflows on a cadence, and a dashboard app over the results.
- **Weekly reporting.** The report format as a skill, a workflow that assembles it from your connected systems every Monday, and the past quarter's reports as files the project can reference.
Each one is a system rather than a task, which is exactly the thing projects are for.
Projects are available in every workspace now. Open **Projects** in the sidebar to create your first one.
## FAQs
### What can I put in a project?
Workflows, apps, and skills, plus files and the chats you have inside it. A resource lives in exactly one project, and you can move existing work into a project at any time.
### Do I have to build everything from scratch inside a project?
No. You can pin skills, workflows, and apps you already built, and they keep working exactly as before. Anything you create from a chat started in the project lands there automatically.
### Are my project chats visible to everyone the project is shared with?
No. Chats in a project are private to the person having them. What the project's specialist knows lives in its memory and its files, which everyone the project is shared with can read.
### What happens to the contents when I delete a project?
You choose. Delete the resources you own along with the project, or keep them in your workspace, where they move to the top level and keep working. Anything someone else owns is never deleted, it moves to your workspace instead.
### Can I share a project with my team?
Yes, and access flows down to what is inside it. Access set directly on a resource still wins, so someone can be an editor on the project and a viewer on one particular workflow in it.
---
# Run the whole thing in your building
> Apple's new desktop holds 512GB of model weights. With a local General Input deployment, prompts, weights, credentials, and runs never leave your network.
Published: 2026-08-27. Last updated: 2026-08-27.
On Tuesday, Apple introduced a new Mac Studio built around a chip called the M5 Ultra ([Apple](https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/)). The number Apple led with wasn't how fast it computes. It was how much memory it holds and how quickly it can read it: up to 512GB, moving at 1.2 terabytes per second. Apple's hardware chief called it "the ultimate desktop for on-device AI."
If you're not deep in this world, that sounds like a spec-sheet detail. It isn't. It's the sign of a change in what a powerful computer is even for.
## The number on the box used to be speed
For about thirty years, you bought a workstation for raw computing power. That made sense, because the work that justified an expensive machine was math-heavy: rendering video, running simulations, compiling code. More calculations per second meant the job finished sooner.
AI doesn't work that way. When a model writes an answer, it produces one word at a time, and for every single word it has to read its entire brain back out of memory. It barely does any math along the way. Almost all of the waiting is reading.
So two numbers decide everything. How much memory you have decides whether a model fits on your machine at all. How fast that memory can be read decides how quickly it answers. Raw computing power is now the third number down the list.
If you're comparing machines for AI, read the memory line first and the speed
line second. A slower chip with more memory can run models that a faster chip
simply cannot load.
## What 512GB actually gets you
Here's a concrete example. DeepSeek publishes a model called V4-Flash, and the file is 167GB. It's free to download and openly licensed.
That size is the whole story. It won't fit on a laptop, or a Mac mini, or even on a single one of the datacenter graphics cards that cost thirty thousand dollars. To run it on smaller hardware you have to compress it first, which means running a rough copy of the model instead of the real thing.
On a 256GB Mac Studio, it loads whole. No compression, no compromise, and no meter running while you use it.
That machine isn't cheap. Configured to the top of what Apple currently sells, it's $18,299, which is more than most people expect a computer without a screen to cost. But it's a thing you buy once, and nobody bills you per question after that.
## A model on your desk is only half the job
Running the model locally answers where the AI thinks. It doesn't answer what the AI is allowed to do.
Real work isn't answering questions in a chat window. It's pulling a record from your CRM, checking a calendar, drafting the follow-up, updating the invoice. The moment AI starts doing that kind of work, it needs access to your accounts, which means your passwords and keys are in play. And if the software coordinating all of that lives in someone else's cloud, then your data is leaving the building anyway. The model stayed home, but the work didn't.
That's the part we built General Input around, and it's why the local model story only recently became a complete one. General Input runs on your own infrastructure. The dashboard, the API, and your database are ordinary services you deploy where you like. When an agent writes and runs code, it does that inside a container on your machines rather than in our cluster. File storage speaks the standard S3 protocol, so a server in your own rack works the same as a cloud bucket.
Because the software is now on your network, pointing it at the Mac Studio is simple. You run a model server on the machine, add its address in settings, and it shows up in the model picker beside everything else. It's just another computer on the same network.
Put that together and the loop closes. The request comes from a server in your building, gets answered by a model in your building, runs code in your building, using credentials that never left your building.
## What still reaches out, and why
One piece doesn't run locally, and it's worth being precise about which.
General Input documents tens of thousands of API operations across hundreds of services, so your agents know how to actually talk to Salesforce or Stripe or NetSuite. That documentation changes constantly as vendors ship updates, and it's identical for every customer, because it describes public APIs. A local copy would be out of date within a week.
So a local install reads that documentation from us as it needs it. That request asks what the Salesforce bulk API expects in a request body. It doesn't carry your Salesforce data.
Ask any vendor promising on-premise AI the same question: what exactly crosses
the network, and what's in it? "Nothing ever leaves" is almost never true. The
useful answer is a specific one.
Your credentials get the same treatment they always have. They're encrypted at rest, injected only when a step needs them, and stripped out before anything reaches the model. Every action taken on your behalf is logged. That holds whether the model is a cloud service or a machine humming under the desk.
## The honest case against doing this today
Two things argue the other way.
Open models still trail the best commercial ones by roughly six to twelve months, and they're weaker on unusual problems that look nothing like their training data. A desktop runs a very good model. It doesn't run the best one, and for some work that gap will matter.
The other is timing. Memory is having its worst pricing year on record. A common 32GB memory kit went from around $120 in late 2025 to nearly $400 this August, and the Mac mini's starting price climbed from $599 to $899 in five months ([Tom's Hardware](https://www.tomshardware.com/desktops/mini-pcs/apple-price-hikes-continue-as-mac-mini-with-16gb-ram-and-256gb-is-now-usd899-1tb-storage-option-adds-usd500-to-entry-level-headless-system)). The capability is heading toward being included with the hardware. The component is getting more expensive on the way there. Both are true, and the second one is a real reason to wait.
## The takeaway
The expensive part of AI was never the model. Models get cheaper and better every few months, and the one you're loyal to today will be replaced.
What's actually hard to replace is everything around it: the connections to your systems, the credentials, the accumulated knowledge of how your business runs. Own that, and which model answers becomes a setting you change on a Tuesday. Own the machine too, and so does the question of whether your work ever leaves the building.
## FAQs
### What actually runs on my infrastructure?
The dashboard, the API server, your Postgres database, the object storage, and the sandboxes that execute agent code. Sandboxes run as local Docker containers rather than in our cluster, and storage speaks the S3 API, so MinIO or any S3-compatible store works.
### What still talks to General Input?
The integration catalog. Operation documentation for tens of thousands of API endpoints is reference data that changes constantly and is identical for every customer, so your deployment reads it from us rather than shipping a stale copy. It describes public APIs, not your data.
### Can the model run locally too?
Yes. Organization custom models point at any OpenAI-compatible endpoint, which is what LM Studio, MLX, llama.cpp, and vLLM serve. When the server is on your network as well, that is just another host on the LAN. No tunnel, no VPN, no public endpoint.
### Does the model ever see my API keys?
No. Credentials are encrypted at rest, injected at runtime inside the sandbox, and stripped before anything returns to the model. That holds whether the model is hosted by a provider or sitting on the desk next to the server.
### What size machine do I need for the model?
It depends on the model, and capacity is the binding constraint. A 27B model at 4-bit is about 17GB and runs on a 24GB Mac mini. DeepSeek V4-Flash is 167GB at full precision and needs a 256GB machine to load uncompressed.
---
# Your AI tools can now reach every system your business runs on
> The General Input MCP connects your workspace to Claude, ChatGPT, Gemini, and Grok, so every integration, workflow, and skill works from the AI tool you use.
Published: 2026-08-25. Last updated: 2026-08-25.
There is a gap between the AI tool you like and the systems your work actually lives in. You can ask Claude a sharp question about a customer, and it has no idea who that customer is, what they pay you, or what they emailed last Tuesday.
The General Input MCP closes that gap. Connect your workspace once, and Claude, ChatGPT, Gemini, or Grok can reach every integration you have authorized, every workflow and skill you have built, and the business context sitting behind them.
## Connectors cover the popular actions. That is the problem.
We measured this before building it. Across the services that appear in both catalogs, a General Input integration carries a median of 35 tools where the equivalent Anthropic-verified connector carries 11. Same vendor, same public API, roughly three times the reach.
The gap is wider than the median suggests on the systems people actually run their week in. Outlook is 46 operations against 3; SharePoint is 33 against 2.
This is not a knock on the people building those connectors. It is a constraint they are working inside. Every tool definition an MCP client loads gets injected into the model's context window before the conversation starts, so shipping the whole API means spending someone's context budget on endpoints they may never call. The rational move is to ship the top ten actions.
General Input takes the other route. Operation docs load on demand and calls execute as code, so nothing occupies context until the agent actually reaches for it. That is why the catalog can be deep without being expensive, and it happens to be the pattern [Anthropic's own engineering team now recommends](https://www.anthropic.com/engineering/code-execution-with-mcp).
## What you can do once it is connected
The connector is not a read-only window onto your data. It exposes the same agent toolset that runs inside General Input, which means from the chat window you can:
- Search and call any operation across your connected integrations
- Build and run workflows, and read their execution traces
- Create skills, then reuse them in later sessions
- Reach the context graph, so the agent knows what your business is and who your customers are
- Write results back into the workspace instead of leaving them in a chat log
Credentials never leave the vault. They are injected at runtime inside the
sandbox and stripped before anything returns to the model, so the AI tool you
connect sees results and never secrets.
## A worked example: the QBR nobody wants to build
Quarterly business reviews are a good test because the work is not the thinking, it is the assembly. CSMs commonly report spending eight to twelve hours putting one together, and the reason is that the numbers live in five systems that do not talk to each other.
In the video above, the ask is one sentence: build a QBR deck for this client, and find what you need in General Input. The agent goes and gets account activity, the history of what was discussed, open issues, and commercial context, then Claude Design lays it out as a deck.
The interesting part comes after. Once the deck exists, you can tell the agent to save the process:
> Now that you understand how to create this deck, please create a skill in General Input with the template and process for this information.
It writes that skill back into your workspace. The next QBR is not another twelve-hour assembly job, it is a repeat of a process your business now owns.
The demo run took twenty to twenty-five minutes of real time across all those
systems. Worth knowing before you try it and expect an instant answer.
## Setting it up
Open **Settings**, find **MCP**, and copy the connector name and server URL. In Claude, go to **Connectors**, choose **Add custom connector**, paste both, and authenticate. That pairs the workspace to your account.
The same server URL works anywhere that speaks MCP. Settings has copy-paste steps for Claude on web and desktop, ChatGPT desktop, Claude Code, and Codex CLI.
## Bring your own intelligence
We have been building toward this for a while. [Bring your own subscription](/models#bring-your-own-subscription) lets you run automations on the ChatGPT, Claude, Gemini, and Grok subscriptions you pay for. The MCP is the other direction: instead of bringing your model to our product, it brings your business to whichever AI tool you already have open.
You should not have to pick a chat window based on which one can see your CRM. Pick the model you like, and connect it to the systems your business actually runs on.
## FAQs
### Which AI tools can connect to the General Input MCP?
Anything that speaks MCP. Claude on the web and desktop, ChatGPT desktop, Claude Code, and Codex CLI all have setup steps in settings, and the same server URL works for Gemini, Grok, or any other MCP client.
### Is the MCP read-only, or can it change things?
It is not read-only. The connector exposes the full agent toolset, so from the chat window you can build workflows, create skills, run workflows, and write results back into your workspace, the same things you would do inside General Input.
### Does connecting the MCP expose my API keys to the AI provider?
No. Credentials stay encrypted in your General Input vault and are injected at runtime inside the sandbox, then stripped before anything returns to the model. The AI tool sees results, never secrets.
### Why does General Input have more tools per integration than a native connector?
Most connectors cover the popular actions of an API. General Input documents the whole surface and loads operation docs on demand instead of injecting every schema up front, so depth costs nothing until a tool is actually used.
---
# Less Than 1% of Candidates Pass This Hiring Bar
> Matt Swulinski grew Wispr Flow to millions of users. He says under 1% of candidates are AI-native systems thinkers, and that's who the market now pays for.
Published: 2026-08-19. Last updated: 2026-08-19.
Matt Swulinski was asked on [20VC](https://www.thetwentyminutevc.com/matt-swulinski) how what he looks for in talent has changed in the last year or two. A year ago, he said, he wanted ten years of experience. Now? **"If they're AI native or a systems thinker and can deconstruct what makes their job hum, that person plus experience will out-compete someone that just has experience."**
And of the candidates he actually meets, fewer than 1% clear that bar.
This isn't a hot take from the sidelines. Matt was marketing hire #1 at Wispr Flow and built its growth function from pre-launch to millions of users. Before that, Superhuman. He scaled Wispr Flow by thinking in systems, so his hiring test is whether you can too.
## What systems thinking actually means
It has nothing to do with which AI tools you've used. It's a loop:
1. Take one step back from the task you're doing.
2. Identify how that task connects to the rest of your job.
3. Ask whether some or all of it can be automated.
4. Automate it, then replace that time with higher-leverage work.
Matt's interview version of this is brutal: ask a candidate what it would take to fully automate their job. Most people, he says, discover they don't really understand their job. They know the pedals and the route. They've never popped the hood.
In the full conversation with Harry Stebbings, he walks through how he built AI-native systems to fuel Wispr Flow's growth. It starts with simple processes, then adds feedback loops that make those processes better over time. That second part is where most people stop short: a prompt with some context is a process. A system critiques its own output and feeds the critique back in.
## The market already prices this
Matt's 1% is one hiring manager's funnel. The aggregate data says the same thing louder.
As far back as May 2024, [71% of leaders told Microsoft and LinkedIn](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part) they'd rather hire a less experienced candidate with AI skills than a more experienced candidate without them. Two-thirds said they wouldn't hire someone without AI skills at all.
Two years later the premium shows up in wages. [PwC's 2026 Global AI Jobs Barometer](https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html), built on a billion job ads across 27 countries, measures a **62% average wage premium for workers with AI skills**, up from 57% the year before. Jobs requiring those skills are growing roughly eight times faster than the market overall.
Experience stopped being the tiebreaker. Legibility of your own job is.
## Why "become AI native" keeps failing
Here's the uncomfortable half of Matt's argument: a lot of companies are trying to make their teams AI-native, and they're failing. His diagnosis is blunt. The people in those roles are not systems thinkers. You can't hand an agent to someone who can't name the parts of their own job.
The failure is measurable. An [MIT NANDA report](https://finance.yahoo.com/news/mit-report-95-generative-ai-105412686.html) found 95% of enterprise gen-AI pilots showed no measurable P&L impact, despite $30-40 billion in spending.
To be fair, it's not only the people. Microsoft's [2026 Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization) finds organizational factors drive 67% of AI's reported impact versus 32% from individual mindset.
Readiness times capability. A team of systems thinkers inside a company that
punishes experimentation stalls just as hard as a ready company staffed with
people who can't name the parts of their own job. Either factor at zero zeroes
the whole product.
This is the same skill gap behind the [agent operator role Aaron Levie sees coming](/blog/agent-operator-million-jobs). The models are good enough. The scarce resource is people who can decompose work, because decomposed work is delegable work.
## There's never been a better time to become AI-native
The vision for General Input is to make these AI-native systems much easier to build: batteries included, low friction, and easy to share across your team. The mapping is still yours to do. Nobody can deconstruct your job but you.
So run Matt's audit on yourself before someone hiring you does. What would it take to fully automate your job? If you can't answer, that's not evidence your job is safe. It's evidence you haven't looked under the hood yet. The 1% who have are getting hired.
## FAQs
### What is an AI-native systems thinker?
Someone who can step back from the task in front of them, map how their whole job fits together, and identify which parts an agent can run. The skill isn't knowing AI tools. It's understanding your own work well enough to delegate pieces of it.
### Does AI skill really beat experience in hiring now?
Increasingly, yes. As far back as 2024, 71% of leaders told Microsoft and LinkedIn they'd rather hire a less experienced candidate with AI skills than a more experienced one without. PwC's 2026 AI Jobs Barometer measures a 62% average wage premium for workers with AI skills.
### How do I become a systems thinker in my current role?
Run the automation audit on yourself: ask what it would take to fully automate your job. Write down every moving piece, the boring admin, the reporting, and how they connect. Then automate one piece and spend the reclaimed time on higher-leverage work. Repeat.
---
# Agents Inside Your Apps: Every Tool You Build Can Now Do Its Own Work
> Embedded agents are live in General Input: any app you build can launch a full agent from its own interface, with the app's connections and record as context.
Published: 2026-08-16. Last updated: 2026-08-16.
An app can show you a ticket. Until this week, it could not do anything about it.
The app in the video above is a support desk for a packaging supplier. It pulls Zendesk tickets, HubSpot records, and rows from a production database into one screen, so a support rep stops cycling through three tabs to understand one customer. That alone is useful. It is also, if we are honest, where most internal tools stop: a nicer window onto work that still has to be done by hand.
Embedded agents are live in General Input as of this week, and they change what that window is for. Any app you build can now carry agents that launch from its own interface. This is not a chat widget bolted to the corner: it is the same full agent that runs your background workflows, started from a button inside the app, with the context of whatever you are looking at.
The video builds one end to end in about fifteen minutes. This post is the longer version: what shipped, why launching an agent from inside an app is different from pasting context into a chat window, and what becomes buildable once your interfaces can do their own work.
## From interface to operating system
The support desk started the video as a read-only unifier. The whole upgrade was one message to the specialist, the agent that builds and edits apps:
> Add an agent to this application, visible on each detail page. It should gather context from all of the systems connected to the application, plus Gmail for past interactions, then give a brief of the issue the customer is facing and draft a response we can send to Zendesk in one click.
GPT-5.6 Sol worked for just under seven minutes. The result is a Draft Reply button on every ticket's detail page. Click it, confirm, and an agent goes off through Zendesk, HubSpot, the production database, and Gmail, then comes back with two things: a brief of what the customer is actually facing, and a drafted response. One more click and the reply is live on the Zendesk ticket.
The second half of the demo matters just as much. The first version of the response interface was not quite right, so the fix was another message: put the response interface in its own tab in the detail view. The app rebuilt around the feedback. When the interface is software you described into existence, iteration is a sentence, not a sprint.
## What the agent inherits
The reason this beats a chat window comes down to what the agent starts out knowing.
**It inherits the app's connections.** The support desk was already authorized against Zendesk, HubSpot, the database, and Gmail. The embedded agent uses the same connections. Nobody re-authenticates anything, and there is no second copy of your credentials living in a bot.
**It inherits the record on screen.** Launch the agent from ticket #27 and it starts with ticket #27. There is no pasting thread history into a prompt or explaining which customer you mean. The context you would have spent five minutes assembling for a chatbot is the starting state.
**It inherits the person.** Every run records who launched it. Share the app with your team and the whole team gets the agent, running through the owner's connections, with no per-seat setup. Live run status streams to everyone in the app, and if an agent hits a question it cannot answer alone, the app shows a Needs input state until a human weighs in.
And because an embedded agent is a real agent rather than a widget, the whole runtime comes with it: full run history in the app's Agents tab, editable instructions, model choice, and a confirmation step before every launch.
## Five patterns worth stealing
The support desk is one instance of a pattern: an interface your team already works in, with agents parked at the exact spots where the work starts. Some versions of it we expect to see a lot of:
- **Support desks that answer.** The demo's shape. A queue, a brief, a drafted reply, one click to send. Service desk benchmarks put an agent-handled ticket at [about $45, three times the cost of a self-service resolution](https://www.mavenagi.com/blog/cost-per-ticket-statistics-customer-support); a drafted reply that starts from full context attacks the expensive part of that number, the twenty minutes of digging before anyone types.
- **A research button in the CRM.** A table of thin lead records, and on each row a button that sends an agent off to fill the record in, summarize what it found, and suggest an angle. Selection is the input; the enriched row is the output.
- **Dashboards that investigate.** An anomaly list where each entry carries an Investigate button. Instead of a chart that tells you something is wrong, an agent that pulls the relevant records and reports back why.
- **A drop box that returns finished documents.** Upload the messy version, get back the house version. The agent does the restructuring; the interface holds the queue and the results.
- **Approval flows with a human at the hinge.** The agent proposes the refund, the reply, or the change, then pauses. A person reviews it inside the app and approves. The agent acts. This is the pattern that makes automation trustworthy enough to touch money and customers.
One app can carry several of these at once, each launched from the place in the interface where its work begins.
This is also the direction enterprise software is already pointed. Gartner expects [a third of enterprise applications to include agentic AI by 2028, up from less than 1% in 2024](https://www.sdxcentral.com/news/according-to-gartner-the-future-isnt-actually-agentic/), with 15% of day-to-day work decisions made autonomously. The open question is who decides which of your tools get agents: your vendors' roadmaps, or you.
## The build-versus-buy collapse
Systems like this are currently a consulting product. Published development guides put custom agent builds at [$20,000 for a simple agent to $500,000 and beyond for multi-agent systems](https://softteco.com/blog/ai-agent-development-cost), with goal-based agents, the kind in the demo, quoted at $40,000 to $150,000. Read the line items in any of those guides and the pattern is consistent: the model is the cheap part. Data preparation, integrations, orchestration, and the interface are the bill.
Not every one of those engagements collapses into a prompt, and it would be dishonest to claim otherwise. But the floor has moved a long way. The support system in the video, interface plus agent plus the wiring between them, took about fifteen minutes on a standard subscription, built by describing it. The version of custom agent software that used to justify a six-figure statement of work is now an afternoon of iterating on your own tool, with the logic of your own business, against systems you already run.
The deeper difference is who gets to change it. A consulting build freezes the day the engagement ends. An app with embedded agents changes whenever someone on your team notices it should: the interface is a conversation away, and so are the agents inside it.
## Live now
Embedded agents are available in every General Input workspace today. Open an app you have built, or build one first, then describe the agent in the chat panel next to it: where it should appear, what context it should gather, what it should produce. The video above shows the whole path in real time.
Interfaces used to be where you kept track of the work. They are becoming where the work gets done.
## FAQs
### Are embedded agents different from regular General Input agents?
No, and that is the point. An embedded agent is the same full agent that powers background workflows: same runtime, same run history, same controls, same billing. The difference is where it launches from and what it starts out knowing. It is declared inside the app, triggered from the interface, and begins with the app's connected systems and the record you launched it from already in context.
### Who can run an embedded agent?
Anyone you share the app with. Runs go through the app owner's connected accounts, so the people you share it with trigger agents without connecting anything themselves, and every run records who launched it. Live run status streams to everyone viewing the app, not just the owner.
### Do I need to rebuild my app to add an agent?
No. Ask for the agent in the chat panel next to the app, the same way you would ask for a new chart or filter. Once it exists, you can edit its instructions or change its model from the app's Agents tab, and the next run picks up the change without a rebuild.
### What stops an agent from doing something I did not want?
Three things. Every launch asks for confirmation before it starts. An agent that reaches a decision above its pay grade pauses and asks instead of guessing, and the app shows a Needs input state until someone answers. And every run is recorded in the app's Agents tab, so you can read exactly what it did and why.
---
# Grok 4.6 hit the Pareto frontier. Don't marry a model provider.
> Grok 4.6 topped CursorBench 3.2 at $2.81 per task while the runner-up charged $17.32. You only capture gains like that if your stack lets you switch models.
Published: 2026-08-13. Last updated: 2026-08-13.
Although Grok started out as an intentionally raunchy alternative to Claude, ChatGPT, and Gemini, Elon and the team at SpaceXAI have turned it into a model at the Pareto frontier.
That is a strange sentence to write about a product whose early differentiator was an unhinged mode. But the benchmarks are the benchmarks.
The Pareto frontier represents the set of optimal choices where you cannot improve one dimension without sacrificing another. Plot every model on a chart with cost on one axis and capability on the other, and most of them land somewhere in the middle of the cloud. There is always something cheaper at the same quality, or better at the same price. The frontier is the outer edge of that cloud, the short list where that is no longer true. To beat a model on the frontier you have to give something up.
Most models never get there. **Grok 4.6 got there by being the cheap one.**
## The cheap one won
[Grok 4.6 was released yesterday](https://x.ai/news/grok-4-6) with the claim that it now sits on that frontier, offering the best combination of cost and intelligence in its class. It jumped past Fable and Sol on [CursorBench](https://cursor.com/cursorbench), which is Cursor's own coding benchmark, built from ambiguous multi-file tasks pulled out of real sessions rather than synthetic puzzles.
The top of that leaderboard is worth sitting with:
- **Grok 4.6** at 70.8%, averaging **$2.81 per task**
- **Fable 5 Max** at 70.5%, averaging **$17.32 per task**
- **Opus 5 Max** at 70.0%, averaging **$8.23 per task**
- **Grok 4.6** at high effort, 69.9%, averaging **$2.34 per task**
Four entries. Nine tenths of a percentage point between the best and the worst of them. And a **7.4x spread in what they cost to run**.
That is not how leaderboards normally look. Usually the expensive model is expensive because it is winning, and the argument is about whether the last few points are worth the premium. Here the model at the top is also nearly the cheapest thing on the board. There is no premium to debate. You are just paying six times more for a slightly lower score.
Cursor doesn't quote headline API rates. It applies each model's published
per-million-token pricing to the tokens that model actually burned on each
task. A model that thrashes gets billed for thrashing.
## The savings aren't a discount, they're fewer steps
That methodology note is the part I find genuinely interesting, and it is buried under the leaderboard where nobody reads it.
If cost per task were just the sticker price, this would be a boring story about xAI undercutting Anthropic on rate card. It isn't. Grok 4.6 finishes the average task in **39 steps and roughly 32,000 tokens**. Opus 5 Max takes 78 steps. Fable 5 Max burns over 100,000 tokens to land six tenths of a point higher.
So the gap is not a discount. The model needs less work to do the work. It flails less, backtracks less, and re-reads the codebase fewer times. On a single task that is a rounding error. Across a team running agents all day, it is most of your bill.
This also quietly reframes what "fast" means. Grok 4.6 is **not** fast in the way people usually mean it. It takes about 32 seconds to produce a first token, and it generates around 86 tokens per second, which is unremarkable. But it gets to done in half the moves. Time to first token is the metric you feel in a chat window. Steps to completion is the metric you pay for in an agent.
## Where the hype needs a haircut
I want to be careful not to oversell this, because the launch coverage already has.
[Artificial Analysis](https://artificialanalysis.ai/models/grok-4-6) runs a broader intelligence index across reasoning, math, and knowledge work rather than coding alone. On theirs, Grok 4.6 comes in **tied for third at 61**, behind Claude Opus 5 at 63 and Fable 5 at 62. Elon called the model "objectively #1." On one benchmark that holds up. On the other it plainly does not.
It is also worth saying out loud that CursorBench is Cursor's benchmark, measuring the kind of work Cursor sells. That is not a knock on the methodology, which is more honest than most, but a coding benchmark is a coding benchmark. If your agents spend their day reconciling invoices rather than refactoring React, the numbers that matter to you may look different.
And Grok is not the cheapest model available. Gemini 3.6 Flash will run you $1.56 a task. It also scores 53.5%, which is a different product for a different job. **The claim worth defending is that Grok 4.6 is the cheapest thing at the frontier, not the cheapest thing.**
## The benchmarks disagree, and that's the argument
Here is the part I keep turning over.
Two credible, well-run benchmarks looked at the same model in the same week and came back with "first place" and "tied for third." Neither is wrong. They are measuring different work.
The instinct is to treat that as a problem to resolve, to find the real ranking. I think that instinct is the mistake. **If the ranking flips depending on which benchmark you trust and what you happen to be doing that day, then picking the right lab was never a strategy.** Not having to pick is the strategy.
That is a much less satisfying conclusion than "use the best model." It is also the only one that survives contact with a leaderboard that reshuffles monthly.
## Intelligence is commoditizing faster than contracts can
This is more evidence that the stream of intelligence is becoming completely commoditized. Competition and innovation are driving prices down while making it easier and easier to switch between models.
Look at the clock on this. Grok 4.6 landed roughly a month after 4.5. It gained five points on the Artificial Analysis index and **did not raise its price**. A 2.1 trillion parameter Grok 4.7 is already announced for a few weeks out. Whatever any leaderboard says today has a shelf life measured in weeks.
Now compare that to the shape of an enterprise AI agreement. Twelve months, sometimes twenty-four, negotiated seats, committed spend. The technology is repricing every few weeks and the contracts are annual. Those two clocks are not close to synchronized, and the gap between them is where your money goes.
Which is why I cannot emphasize this enough: if you marry a single model provider and commit exclusively to its models and products, you will miss out on these gains in cheaper, more capable intelligence.
## Your ceiling is set by someone else's roadmap
ChatGPT serves OpenAI models. Claude serves Anthropic models. That is not a flaw, it is the definition of the product, and both are excellent at what they do.
But it does mean the ceiling on the intelligence available to you is set by one lab's roadmap and one lab's release schedule. Everything above is a thing that happened to other people if your work lives inside a first-party app.
The half-measures deserve more scrutiny than they get. Microsoft 365 Copilot is the flagship "we support model choice" product and genuinely earned credit for [breaking OpenAI exclusivity](https://www.microsoft.com/en-us/microsoft-365/blog/2025/09/24/expanding-model-choice-in-microsoft-365-copilot/). Then look at the actual picker. [Five options](https://learn.microsoft.com/en-us/microsoft-365/copilot/cowork/cowork-models): Auto, Claude Opus 4.8, Claude Sonnet 4.6, an advisor mode, and GPT 5.5. No Grok. And it still lists Opus 4.8 rather than Opus 5.
Model choice, as a feature, means a shortlist someone else curates on someone else's schedule. **That is a different thing from being able to point at anything the day it ships**, and the distinction only shows up on days like yesterday.
OpenAI, Google, and Anthropic have no incentive to give you access to better-value agentic options once they have locked you into an annual enterprise contract. Nobody at any of those companies is going to add a competitor to your dropdown the week that competitor starts winning.
## Own the platform, rent the intelligence
Owning your platform and use cases, while treating models as interchangeable infrastructure, will pay off in the long run.
For what it is worth, we route a lot of our own work to Grok 4.6. It is the primary model in two of our three automatic routing tiers, it was routable here the day it shipped, and selections that were pinned to Grok 4.5 upgraded themselves. Nobody migrated anything. When 4.7 lands, that will be a config change rather than a project.
None of that is a bet on Grok. It is a bet against having to bet. Three months ago the same slot was filled by a different model, and three months from now it probably will be again. The routing survives the churn because the model is a setting rather than a foundation.
**The durable advantage was never picking the right model. It is being able to change your mind cheaply.**
Can't wait to see this trend continue.
---
# Meta open-sources Muse Spark 1.2
> Meta is releasing open weights for Muse Spark 1.2, a near-frontier model at $1.25/$4.25 per million tokens. What it means for prices, hosts, and open source.
Published: 2026-08-10. Last updated: 2026-08-10.
[Meta announced this week](https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html) that it will be open-sourcing Muse Spark 1.2, a near-frontier-level model led by Alexandr Wang.
The reversal is what makes this interesting. Meta launched the Muse line closed back in July and took real heat for it -- the eulogies for American open source were already half-written. One month later, the weights are on the way. And alongside the Spark announcement, [Muse Glimmer](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) is already out: a 30B agentic model under Apache 2.0 that quantizes to under 20 GB and runs on a single consumer GPU. Glimmer's weights are live today; Spark 1.2's are due in the coming weeks.
The model is currently on the cheaper end of the spectrum, at [$1.25/million input tokens and $4.25/million output tokens](https://artificialanalysis.ai/articles/muse-spark-1-2), compared with Fable 5 at $10/$50, respectively.
It’s also extremely fast, averaging 121 TPS on Meta’s infrastructure.
To be clear about where it sits on capability: [Artificial Analysis scores it 54](https://artificialanalysis.ai/articles/muse-spark-1-2) on their Intelligence Index, dead even with Grok 4.5 and seven points behind Claude Opus 5. Seven points is a real gap. Most workloads will never notice it, though, and near-frontier intelligence at roughly a tenth of the blended token price is an enormous market. AA's own framing is that Spark 1.2 is [among the most cost-efficient models at its intelligence level](https://x.com/ArtificialAnlys/status/2085116738023399710) -- about $0.40 of spend per task on their index, versus $1.18 for GPT-5.5.
Generally, I have two thoughts:
1. Models are becoming completely commoditized. [Fireworks and Together will host this model](https://x.com/alexandr_wang/status/2086756152034066792) for the same price or cheaper and let organizations fine-tune it to reach frontier-level intelligence for their specific use cases. Glimmer was live on Together, Fireworks, OpenRouter, Ollama, and vLLM the day it dropped, and Spark 1.2 will land the same way. When every host is serving identical weights, nobody can charge for the model anymore -- only for the serving: throughput, latency, uptime, and price per token. That competition grinds serving margins down toward raw GPU cost, and it only exists because the weights are public. This should generally scare the hell out of Anthropic and OpenAI, as there will be increasing price pressure heading into their IPOs.
2. Having this alternative in the market will decrease reliance on Chinese open-source models and lessen the pressure everyone is feeling to regulate open-source models in the U.S. Since Llama stalled, the open-weights race has basically had one side. [Aaron Levie put it plainly](https://x.com/levie/status/2086802472950239618): "America now finally has its response to the open weights AI race." Now that there's a big dog in the fight on the American tech side, it's going to be much less likely that frontier-level open source gets banned outright.
The fine-tuning piece deserves more attention than it's getting. Glimmer ships with [TorchTitan recipes](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) out of the box, and Together already runs LoRA fine-tunes on the open models it hosts. You can teach a near-frontier model your specific accounts and edge cases, then run it anywhere you want -- including on hardware you own. A closed API can't really offer that. Everyone rents the same generic model, at whatever price the vendor decides intelligence is worth this quarter.
As I’ve been saying, I’m excited to see intelligence trend toward “too cheap to meter” and increase the number of things we’re able to build with an essentially unlimited stream of intelligence.
---
# General Input Crash Course: Sign Up, Connect, Build a Workflow, Ship an App
> A sixteen-minute crash course that goes from the sign-up page to a shipped system without cutting away from the hard parts.
Published: 2026-08-09. Last updated: 2026-08-09.
Most product tours show you features. This one shows a build.
We recorded the whole path in, from the sign-up page to a system that works, and cut it into a single crash course. Sixteen minutes, no cutaways at the hard parts. Everything in it is real: real Google accounts, a real form submission, real internet research, and an app at the end that is reading the rows the workflow wrote a few minutes earlier.
## What actually gets built
The scenario is a staffing conference. It is a good demo because it is small enough to follow and real enough to break.
A QR code at the booth points at a Google Form. Submissions land in a Google Sheet. Every time someone fills the form in, an agentic workflow researches who they are, enriches their row in the sheet, and emails a short brief. After the conference, an app reads all of it back: headline metrics, an analytics tab, a list of every attendee, a detail page per person, and a button that drafts a follow-up email straight into Gmail.
By the end there are three durable things in the workspace: a skill that carries the writing voice, a workflow that enriches every lead as it arrives, and an app to work the results. None of them were configured in a settings page. All three were described out loud.
## The seven steps
- **Sign up and set up the workspace** (00:40). Naming the workspace, picking a plan, and the two onboarding questions that matter more than the thirty seconds they take: what your business does, and what you personally work on. Everything an agent later writes on your behalf reads better for it.
- **Connect your business systems** (02:48). Google Forms by one-click OAuth, then Exa for internet research by API key. Anything you connect is immediately available to every agent, workflow, and app. You do not wire it up again per automation.
- **Choose your models** (04:01). How model selection works, what the Auto option does, and how existing AI subscriptions can power automations. Subscription setup now lives in Subscription Models inside General Input. [Models](/models#bring-your-own-subscription)
- **Build a skill** (06:16). A skill built by reading the emails you have already sent, rather than by writing a style guide from scratch. It works out how you write and who you write to, then gets refined by chat with copy from the website.
- **Chat with your systems** (09:26). The conference setup, described in one go instead of built one piece at a time.
- **Build the workflow** (11:06). Form submission triggers research, the sheet row fills in, and a brief lands in the inbox. Then a real form submission goes through it on camera.
- **Build the app** (13:20). Fifty-one leads in a spreadsheet turned into an interactive tool, described in the same chat thread that had been running all along, so it already had the context.
## Three moments worth skipping to
**[10:09](https://youtu.be/DAb3ju9Patg?t=609), one message becomes three things.** Rather than creating a form, then a sheet, then a QR code, the whole system gets rambled into a single message. Three artifacts come back, built across live authorized accounts rather than a copy of your data. It also pulls in the skill from the previous section without being asked twice.
**[11:46](https://youtu.be/DAb3ju9Patg?t=706), the workflow tests itself.** After the workflow is created it runs itself end to end with realistic data, hits a problem, fixes the workflow, republishes, and tests again. This is the part we would point at if we could only point at one thing. You are not guessing whether the first real submission will work, and you are not debugging in production with a live lead as the test case.
**[14:34](https://youtu.be/DAb3ju9Patg?t=874), the app is just a conversation.** The finished app has metrics, charts, a list view, detail pages, and a Gmail draft flow. It took about seven minutes to build. Restyling it at the end is not a rebuild, it is a message in the chat next to it, and the app updates live.
## What you should take from it
Four things generalize past the conference demo.
**Connect once, use everywhere.** Integrations are authorized at the workspace level, not per automation. The Gmail connection made for the skill is the same one the app drafts through an hour later.
**Context is a resource, not a preamble.** The skill built at 06:16 is used by the chat at 10:09, the workflow at 11:06, and the app at 13:48. Tone and audience stay consistent because they live in one place instead of being re-explained per task.
**Describing beats configuring.** Every artifact here was created by talking. The output is still inspectable afterwards: workflows are plain-text readable, every variable can be swapped, and executions stream in one at a time and can be watched live.
**Setup is not a separate chore.** The same chat that answers questions about your data can go and create the things you are describing, because it is working against the accounts you authorized rather than a copy of them.
The demo system in the video is a lead tracker because a lead tracker is easy to follow in sixteen minutes. The interesting version is the one where the systems are yours and the domain knowledge is yours. That part takes about the same amount of time.
## FAQs
### Is the demo real or is it edited to look fast?
It is real. Real Google accounts, a real form submission, real internet research through Exa, and the app at the end is reading the rows the workflow wrote. The one thing compressed for time is waiting: the app took about seven minutes to build and you do not sit through all of it.
### Do I need to watch all sixteen minutes to get value?
No. The crash course is chaptered. If you only have two minutes, jump to 10:09, where one spoken message creates a Google Form, a Google Sheet, and a QR code across live accounts.
### Do I need to know how to code?
No. Every artifact in the video is created by describing it out loud. Workflows are plain-text readable afterwards, so you can see exactly what will run and change any variable, but nothing in the build requires writing code.
### What does it cost to follow along?
You can start free. The video picks the Team plan because it allows up to ten users and includes enough credits to build something real rather than a toy. You can also connect your existing ChatGPT or Claude subscription from Subscription Models in General Input.
---
# How to migrate your Relay.app workflows: a step-by-step guide
> Relay.app shuts down August 15, 2026 for free users and September 14 for paid plans. How to export each workflow, rebuild it from the AI export, and cut over.
Published: 2026-07-21. Last updated: 2026-07-21.
Relay.app [announced on July 16](https://www.relay.app/) that it's shutting down. Free accounts close August 15, paid accounts September 14, and after those dates the data is gone. The Relay team is handling the wind-down well, with prorated refunds, 60 days of free access for paid workspaces, and full exports. But the dates are firm.
- **Now:** exports are available, and new signups and upgrades are off
- **August 15, 2026, 11:59 PM PT:** free accounts and their data are deleted
- **September 14, 2026, 11:59 PM PT:** paid accounts and their data are deleted (access is free until then)
This guide moves one workflow from Relay to General Input, start to finish. Pick one that matters but won't hurt anything while you test. Once you've done one, the rest go the same way, and there's a note on that at the end. Plan on 15 to 30 minutes for the first one.
## Step 1: Create a General Input workspace
Sign up at [generalinput.com](https://generalinput.com) and complete the onboarding. You'll pick a plan along the way, free or paid. If you go paid, the code **RELAYAPP** takes 50% off your first month at checkout.
## Step 2: Hit New in the Workflows tab
Open the Workflows tab and hit **New**. Select **Agentic Task**, then hit **Start from scratch** in the lower left corner.
## Step 3: Copy the build prompt from Relay.app
In Relay, open the workflow, hit the three-dot menu in the top right, and choose **Copy build prompt**. That puts a plain-language description of what the workflow does on your clipboard, and it's everything you'll need to rebuild it.
While you're in Relay, export your run history and tables as CSV too, even for workflows you're not touching yet. You can't export anything once your account's gone.
## Step 4: Paste in the build prompt
Paste the build prompt into the chat where it says "Tell me what you'd like to automate," and send it. That description is all the agent needs.
## Step 5: Watch your workflow get built
The agent rebuilds the workflow from the description: the trigger, the integration steps, and the approval steps where Relay had them. Approvals are a normal step in General Input, so a workflow that waits for a manager's sign-off keeps working that way.
## Step 6: Chat with the agent until you're happy
Read the draft against the Relay original and tell the agent what to fix, like a schedule or a filter condition that drifted. If the workflow has approval steps, double-check who gets asked and when. When it asks for the accounts the workflow uses, connect each one. Your keys are encrypted and never shown to the AI, and every call the workflow makes is logged.
When it looks right, enable the workflow. Then iterate: watch the first few runs and keep telling the agent what to tweak until it's exactly how you want it.
## Then do the rest
Run the new workflow next to the Relay original for a few real triggers and compare the results. When you trust it, turn the Relay one off and disconnect its app connections in Relay.
Repeat steps 2 through 6 for each remaining workflow. Start with the ones that'd hurt most if they broke, so they get the most testing runway. Free accounts have until August 15, paid accounts until September 14, so don't leave the important ones for the last week.
The build prompt is what makes this mostly paste and review. It describes what each workflow does in plain language, and that's exactly what General Input builds from.
## FAQs
### When is Relay.app shutting down?
Relay.app announced on July 16, 2026 that it is winding down. Free accounts and their data are deleted after August 15, 2026 at 23:59 PT. Paying customers keep full access, free of charge, through September 14, 2026 at 23:59 PT, with automatic prorated refunds for annual plans and bonus steps and AI credits during the transition.
### How do I export my workflows from Relay.app?
From your Relay.app workspace you can export workflows, sequences, and MCP servers as JSON and as AI prompts, plus run history and tables as CSV files. For migrating, the AI-prompt export is the one you need. It is a plain-language description of each workflow that you can use to rebuild it on another platform.
### What is the best Relay.app alternative for human-in-the-loop workflows?
Look for a platform that treats approvals as a normal step rather than a workaround. Relay.app made human-in-the-loop a first-class feature. General Input works the same way: a workflow can pause, ask a person to approve or edit something, and then continue. Approval-heavy Relay workflows move over intact instead of being split in two and glued together with email links.
### How do I migrate Relay.app workflows to General Input?
Export each workflow from Relay.app as an AI prompt, then paste that prompt into General Input. The agent rebuilds the workflow with its trigger, integrations, and approval steps, and waits for your review before the first run. Reconnect your accounts, run it alongside the Relay original while Relay is still up, and cut over once you trust it. The code RELAYAPP takes 50% off your first month.
---
# Overpaying for incremental intelligence
> The best AI model is often a few percent smarter than a rival and several times the price. That premium caps how much intelligence you can afford to run.
Published: 2026-07-20. Last updated: 2026-07-20.
Anthropic's Claude Fable 5 is 11% more intelligent than xAI's Grok 4.5 on the Artificial Analysis Index. It is also 567% more expensive. The chart above puts every major lab's flagship on those same two axes, and it tells one story: intelligence has bunched up, and price has not.
## The releases stopped being about the frontier
The last two weeks made that hard to ignore. Meta shipped Muse Spark, xAI shipped Grok 4.5, and Moonshot shipped the open-weight Kimi K3, which landed at 57 on the index, effectively tied with the best models OpenAI and Anthropic sell ([Artificial Analysis](https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5)). The six leading flagships now sit inside a ten-point band, 50 to 60. Ten points separate the best model in the world from the sixth-best, and on most real work you would not notice the difference.
## The premium buys almost nothing
Now look at what the premium actually buys. Blended at a standard 3:1 input-to-output mix on OpenRouter list rates, prices run from about $2 per million tokens for Muse Spark to $20 for Fable 5. Moving from Kimi K3 to Fable 5 buys you three index points and roughly triples the bill. Grok to Fable 5 buys six points at 6.7x the price. This is a diminishing-returns curve, and the habit of defaulting to the best model puts you at the steep, expensive end of it, paying a luxury markup for intelligence most tasks never call on.
## The real cost isn't the invoice, it's the throttle
Here is the part that never shows up on a per-token price sheet. Your intelligence budget is finite. A model that costs 6.7x as much doesn't just cost more per call, it means 6.7x fewer calls for the same money. The premium quietly sets a ceiling on how much intelligence you can run at all.
That ceiling is where the real money leaks. Picture an agent that watches customer activity every hour, flags the accounts showing buying intent, and drafts outreach the moment a signal appears. On the cheap tier you run it continuously. On the frontier tier, at 6.7x the cost, the same budget only buys a bulk pass once a week. By then the prospect has already bought from someone else. You didn't pay for a smarter answer. You paid for a slower one.
As companies push programmatic intelligence into every workflow, this inverts the question. The constraint stops being "which model is smartest" and becomes "how much intelligence can I afford to run around the clock." Price per token is the new bottleneck.
## The ROI question the market is now asking
This lands at an awkward moment for the premium labs. OpenAI and Anthropic both filed IPO prospectuses in June, targeting roughly $1 trillion and $965 billion respectively ([Barchart](https://www.barchart.com/story/news/3116486/openai-anthropic-ipo-what-s-confirmed-what-s-still-speculation)). Those numbers assume the frontier stays a moat you can charge a steep premium for. But the underlying economics are strained: one analysis pegs the three headline AI IPOs as needing on the order of $150 billion in combined annual earnings to justify their valuations, against roughly $20 billion in combined losses today ([Intellectia](https://intellectia.ai/blog/ai-mega-ipo-wave-2026-spacex-openai-anthropic)). If a $6 open-weight model is 95% as capable as a $20 one, the premium tier's pricing power is precisely the assumption under the most pressure.
## Make intelligence a dial
The reason teams overpay is rarely ignorance. It's friction. In most stacks the model is welded in: prompts tuned to one provider, parsing built around one format, the whole thing tested against a single API. Switching is a migration nobody wants, so the premium compounds one request at a time.
General Input removes the friction. The model is a setting, chosen per workflow. Run a customer-facing answer on the frontier and a bulk classification job on a mid-tier model. Or hand the decision to a router that sends each task to the cheapest model clearing your quality bar, and re-checks as cheaper models keep improving. Because the model never sees your credentials, which are encrypted at rest, injected at runtime, and stripped before anything reaches it, with every external call logged, swapping models is a config change, not a security review.
The winners in the next phase of this won't be whoever owns the single smartest model. They'll be whoever can deploy the most intelligence per dollar. You're paying frontier prices for a rounding error of extra intelligence. Price the increment before you buy it.
## FAQs
### Is the most intelligent AI model worth the price?
Usually not, for most tasks. The six leading flagships now sit within ten points of each other on the Artificial Analysis Intelligence Index (50 to 60), while their blended prices span 10x, from about $2 to $20 per million tokens. Moving from a near-frontier model to the top one often triples or 10x-es the bill for a few index points that most workloads never use. The frontier model is worth it for the specific tasks that genuinely need it, not as a default for every call.
### How much cheaper are non-frontier models?
A lot. Blended at a standard 3:1 input-to-output mix on OpenRouter list rates, Meta Muse Spark runs about $2 per million tokens, Grok 4.5 about $3, Kimi K3 about $6, and GPT-5.6 Sol about $11, against $20 for Claude Fable 5. Kimi K3, an open-weight model, scores 57 on the intelligence index versus Fable 5 at 60, so you keep roughly 95% of the capability at under a third of the price.
### What is the efficient frontier for AI models?
It is the set of models where you cannot get smarter without paying more, or pay less without getting less capable. A model that another model beats on both price and intelligence is dominated and sits off the frontier. Choosing well means picking the cheapest model on the frontier that still clears the quality bar for your specific task, rather than defaulting to the most capable model for everything.
### How does General Input reduce model costs?
It makes the model a per-workflow setting instead of a hardcoded foundation. You can route a customer-facing answer through a frontier model and a bulk job through a cheaper one, or hand the choice to a router that sends each task to the cheapest model clearing your quality bar and re-checks as cheaper models improve. Because the model never sees your credentials and every external call is logged, swapping models is a config change, not a security review.
---
# Talking to AI agents is 3x faster than typing to them
> Voice dictation is live in General Input. Why speech is 3x faster than typing, and how AssemblyAI plus a Gemini cleanup pass turns rambling into a prompt.
Published: 2026-07-08. Last updated: 2026-07-08.
Here's a raw transcript from the demo above:
"Um, so, take my notes from the, the Q2 pipeline review and uh... turn them into like a one-pager, a branded PDF, for the board meeting tomorrow."
And here's what General Input put in the prompt box:
"Turn the Q2 pipeline review notes into a branded one-page PDF for tomorrow's board meeting."
Voice input has been one of the biggest feature requests since we launched: voice to workflows, voice to apps, voice to queries. It's live today. Tap the mic in the prompt box, talk, and the prompt writes itself. Here's why that's worth caring about, and how the cleanup actually works.
## Where the speed comes from
The average person types [around 40 words a minute](https://en.wikipedia.org/wiki/Words_per_minute). Conversational speech runs 120 to 150. That gap exists because typing makes you do two jobs at once: compose the thought, then serialize it through your fingers, one keystroke at a time, backspacing over typos as you go. Speech skips the second job entirely. Your thoughts already come out as spoken language; there's no extra motor step between having the sentence and producing it.
This isn't a new observation. In 2016, researchers at Stanford, the University of Washington, and Baidu ran the head-to-head: speech input was [3.0x faster than the keyboard for English, with a 20.4% lower error rate](https://hci.stanford.edu/research/speech/). The researchers [expected a slight edge and got a blowout](https://engineering.stanford.edu/news/smartphone-speech-recognition-faster-and-more-accurate-typing).
Prompts are the worst case for typing. An agent does its best work when you give it context, constraints, and what the output should look like. That's a paragraph or three, many times a day. It's exactly the kind of text where a 3x difference compounds, which is how a dictation app like Wispr Flow ends up [reportedly raising at a $2 billion valuation](https://www.bloomberg.com/news/articles/2026-05-12/ai-dictation-startup-wispr-in-funding-talks-at-2-billion-value).
## Why dictation never stuck
Read the raw transcript at the top again. It has filler, a repeated word, a trailing "uh", and a correction mid-sentence. That's how everyone talks. Older dictation tools transcribed it faithfully, so you either sent embarrassing text or you dictated like a court stenographer, speaking punctuation out loud in clean, pre-planned sentences. Planning your sentences before saying them burns the entire speed advantage. Transcription accuracy got solved years ago; the ramble problem is what kept keyboards in charge.
## What happens after you stop talking
Two systems run back to back.
While you talk, the browser streams your audio directly to AssemblyAI over a short-lived token our server mints for the session. The audio never passes through General Input's servers. Words appear as you say them.
When you stop, the finished transcript and whatever was already in the prompt box go to Gemini with one job: produce the text you meant. Concretely, it:
- Drops filler and false starts, and resolves self-corrections down to the corrected version. "The, the Q2 pipeline review" becomes "the Q2 pipeline review"; "make it for Thursday, actually Friday" becomes "Friday."
- Fixes punctuation, casing, and obvious mis-transcriptions.
- Applies spoken instructions instead of transcribing them. Say "delete that last sentence" or "make it shorter" and it edits the draft rather than writing the words "delete that last sentence" into your prompt.
- Weaves the new speech into whatever you'd already typed, and answers in the language you spoke.
That third behavior is the part we like most. It means the mic isn't just dictation, it's editing. You can draft a prompt by voice, look at it, and keep talking to revise it.
And if the cleanup model is ever unavailable, the raw transcript still lands in the box. You lose the polish, never the words.
## Closer to the speed of thought
Building on General Input was already 10x faster than assembling the same automation in Zapier, n8n, or Make. The keyboard was the next slowest link. Prompts are the highest-leverage text you write all day. Stop typing them.
## FAQs
### How does the cleanup actually work?
Two systems run back to back. While you talk, the browser streams audio to AssemblyAI and words appear live. When you stop, the finished transcript plus whatever was already in the prompt box goes to Gemini, which drops filler and false starts, resolves self-corrections to the corrected version, fixes punctuation and obvious mis-transcriptions, and merges the result into your draft.
### Does my audio pass through General Input servers?
No. Your browser streams audio directly to AssemblyAI using a short-lived session token that General Input mints for you. The audio itself never touches our servers; we only see the finished transcript when it comes back to be cleaned up.
### Can I edit what I already wrote by voice?
Yes. Spoken instructions get applied instead of transcribed. Say "delete that last sentence" or "make it shorter" and the cleanup pass edits your draft rather than writing those words into the prompt. You can draft by voice, look at it, and keep talking to revise.
### What happens if the cleanup model is unavailable?
The raw transcript still lands in the prompt box. You lose the polish, never the words.
---
# Why Running Chinese Open-Source Models Yourself Is the Safest and Cheapest Option
> Why self-hosted open-weight models like GLM, Kimi, and DeepSeek are both cheaper and safer than closed frontier APIs, explained down to the mechanism.
Published: 2026-07-06. Last updated: 2026-07-06.
The exchange keeps happening in enterprise AI, almost word for word. "We can't use an open-source model. It's all Chinese." Even if it's 100x cheaper? "No. We care about safety and security."
That answer gets both halves of the trade-off backwards. Running open-source models like GLM, Kimi, and DeepSeek yourself is usually cheaper and safer than piping your data to a closed frontier API. Not as a slogan, as a matter of how the two things actually work. Start with the one fact everything else follows from.
A model is a file. A few hundred gigabytes of numbers that, given some text, predict the next token. Anthropic and OpenAI keep their files secret and run them only on their own servers. The labs behind GLM, Kimi, and DeepSeek publish theirs, so anyone can download the file and run it. That single difference, secret-and-hosted versus published-and-yours, is what makes one path expensive and exposed and the other cheap and private.
## Why it's cheaper
When you call a closed API, you're paying for three things bundled together: the compute to run the model, the cost of training it, and the lab's margin. There's exactly one seller of that specific model, so there's no price competition on it. You pay what they charge.
An open model unbundles that. The weights are free, so nobody is charging you to recoup a training run or protect a franchise. You pay for compute, and only compute. Then two mechanisms drive that compute cost down hard.
The first is architecture. Modern open models use a mixture-of-experts design, where only a small fraction of the network activates for any given token. The file is huge, but the work per token is small, which is [why NVIDIA measures roughly a 10x drop in cost per token](https://blogs.nvidia.com/blog/mixture-of-experts-frontier-models/) versus dense models of similar quality.
The second is competition. Because the weights are public, dozens of providers, Fireworks, Together, Groq, DeepInfra, all serve the exact same model and compete on nothing but price and speed. That races the cost toward the raw price of the GPUs. [DeepInfra drove one large open model from 20 cents per million tokens to 10 to 5](https://blogs.nvidia.com/blog/inference-open-source-models-blackwell-reduce-cost-per-token/) just by moving hardware generations. [NBC News reports](https://www.nbcnews.com/tech/innovation/silicon-valley-building-free-chinese-ai-rcna242430) the net effect is Chinese open models running ten to fifty times cheaper than Western equivalents. Nobody can do that to a closed model, because nobody else is allowed to serve it.
## Why it's safer
Now the part people get exactly wrong. Safety was never a property of the model. It's a property of where the model runs.
A closed model is a service. Every prompt you send, the contract, the customer record, the roadmap, crosses the wire to the lab's servers, and it lives there. "Delete" is a policy, not a physical fact. The clearest proof arrived in 2025, when a court ordered OpenAI to [preserve all ChatGPT and API output logs indefinitely](https://www.terms.law/2025/11/12/openai-v-new-york-times-stopped-being-just-a-copyright-case-the-moment-the-court-turned-to-your-chatgpt-logs/), overriding its own deletion policy and its users' delete requests, and later to hand 20 million de-identified conversations to opposing counsel. Once your data is on someone else's servers, its fate is out of your hands. A subpoena, a breach, or a policy change reaches it, and you find out afterward.
A published model has none of that. You download the file and run it on your own hardware, or in a US datacenter you control on AWS, Azure, or GCP. The weights have no channel back to their creator. As Interconnects' Nathan Lambert [puts it](https://www.interconnects.ai/p/what-people-get-wrong-about-the-leading), "model weights alone cannot reveal anything to their creators." Beijing cannot see a prompt that never left your VPC, and neither can anyone else.
Hosted on a Chinese company's own API, a Chinese model does send your prompts
across the Pacific, and nobody should route sensitive data there. Self-hosted,
the country of origin is a fact about the training run, not about where your
data goes.
There's a second, subtler exposure the "safe" default carries. A frontier lab isn't just a vendor, it's a company with its own product roadmap, and lately it points at its customers. Figma and Cursor both built on Anthropic, then watched it launch [Claude Design into Figma's market](https://venturebeat.com/technology/anthropic-just-launched-claude-design-an-ai-tool-that-turns-prompts-into-prototypes-and-challenges-figma), knocking [7% off Figma's value in a day](https://www.fastcompany.com/91538439/design-enters-its-frenemies-era), and Claude Code into Cursor's. Nobody has to read your individual prompts to benefit from that arrangement; aggregate usage alone maps which categories are worth entering. Self-hosted weights don't have a product team, and they can't grow into your competitor.
## The risks that survive the debunk
Open weights aren't risk-free. [CEPA is right](https://cepa.org/article/the-ai-kill-switch-dangerous-chinese-open-source/) that backdoors buried in weights are hard to detect, and censorship on China-sensitive topics is baked in. But notice these are output risks, and they apply to every model from every lab. Any model can be prompt-injected or emit a bad dependency. Trusting a particular lab doesn't fix that. Architecture does: least privilege, secrets kept out of the model's reach, sandboxed outputs, everything logged.
## Trust the architecture, not the flag
If your system is only safe when the model is benevolent, it isn't safe. That's the principle General Input is architected on: the AI never sees an API key. Credentials are encrypted at rest, injected at runtime, and stripped before anything reaches the model, and every external call it makes lands in an audit trail. The trust boundary sits below the model, so provenance stops being a security review and becomes a line item, price and latency and benchmark score.
The model doesn't matter. The host does. And the safest, cheapest host is you.
## FAQs
### Can a Chinese open-source model send my data back to China?
Not when you self-host it. The model is a file of weights with no network access of its own. Run it on your own hardware or in a US datacenter you control, and your prompts never leave that boundary. The country of origin describes where the model was trained, not where your data goes. Sending prompts to a Chinese-hosted API is a different story, and that data does leave.
### Why are open-source models so much cheaper than closed APIs?
The weights are free, so you only pay for compute. Mixture-of-experts models activate a small fraction of the network per token, and because dozens of inference providers serve the exact same open model, they compete purely on price. A closed model has one seller and no price competition, so you pay for compute plus training recovery plus margin.
### Is a self-hosted open model really safer than OpenAI or Anthropic?
For data privacy, yes. A closed model runs only on the lab's servers, so every prompt lives there. A 2025 court order forced OpenAI to retain all logs indefinitely, overriding its own delete policy. A self-hosted file has no channel back to anyone, so there is no counterparty to trust in the first place.
### What are the real risks of open-weight models then?
Backdoors hidden in weights and baked-in censorship on China-sensitive topics. But these are output risks that apply to every model from every lab, and the mitigation is architectural: least privilege, secrets kept out of the model, sandboxed outputs, and full logging.
---
# The AI use case I keep seeing: documents in bulk from unstructured data
> Turning messy inputs into formatted documents is AI's most repeatable win in operations. How to move it from your laptop to the cloud for your whole team.
Published: 2026-07-02. Last updated: 2026-07-02.
Across every team we work with, one use case keeps showing up, and it's a bigger time saver than almost anything else: building and formatting documents in bulk from unstructured data.
The shape is always the same. The facts arrive messy: a run-on Slack message, a forwarded email, a CSV export, a folder of photos. The output has to be exact: a branded PDF or DOCX with the right fonts, the right layout, the boilerplate that isn't allowed to be wrong. Proposals, client one-pagers, reports, onboarding packets. Nothing needs to be written. Everything needs to be placed.
## The cleanest example: the listing one-sheet
Real estate makes the pattern easy to see. The moment a listing goes live, every fact about it already exists: four beds, three baths, 2,450 square feet, $739,000, open house Saturday 1 to 3, twenty-four professional photos. It's all sitting in the MLS, complete and correct.
And yet someone is about to spend an hour in Canva turning those facts into a one-sheet. The thinking portion of that hour is zero. The price is the price. One hundred percent of the labor is presentation, which means one hundred percent of it is automatable without asking an AI to invent a single word.
The volume hides in the mutations. Each listing regenerates its collateral at every status change: just listed, open house, price improvement, pending, sold. Multiply across a brokerage of forty agents and the marketing coordinator is running hundreds of identical conversions a month, each urgent, each by hand.
## Yes, Claude can do this on your laptop
This isn't a hypothetical capability. Drop the export in a folder, hand Claude your template, and it will produce the document, correctly, today. If you're one person with a recurring formatting chore, you should be doing exactly this already.
But watch what happens when a second person needs the same document. Their prompt is a little different. Their copy of the template is three versions old. They forgot the compliance footer because the compliance footer lived in their coworker's head. The output is close, and close is exactly the problem, because brand and boilerplate are the parts of a document where "close" gets you an email from legal.
A local process is a personal trick. It saves one person's evening and leaves everyone else's untouched.
## Deploy it once, and the whole team inherits it
This is why the process belongs in the cloud. On General Input, the same conversion runs as a workflow the whole team shares. The intake sits where the facts already show up: forward the email, or drop the message in the team's Slack channel. The agent reads the facts, pulls the photos from the shared folder, and renders the collateral with the organization's actual fonts, actual layout rules, and the footer included, every time, because the rules live in the workflow instead of in whoever happens to be formatting tonight.
Consistency stops being a training problem. Fifty people get identical output from day one. When the brand refreshes or the boilerplate changes, you update the template once and everyone's next document is already right. Nobody redistributes a file, nobody re-explains the rules, nobody quietly keeps using the old version.
Nothing was generated. Every number traces back to the source. The AI's entire job was the hour nobody wanted: the conversion. The difference between doing that trick locally and deploying it is the difference between saving yourself an evening and deleting the chore from your whole team's calendar.
## FAQs
### What kinds of documents does this work for?
Anything where the facts already exist and only the formatting is missing: listing one-sheets, client one-pagers, proposals assembled from call notes, recurring reports, onboarding packets. If the job is placing known information into a fixed template, it converts cleanly.
### Can Claude really do this on my laptop?
Yes. Drop the export in a folder, hand it your template, and it will produce the document correctly today. That's the right move for one person with a recurring formatting chore. It breaks down when a second person needs the same document, because prompts, template versions, and compliance details drift from person to person.
### What changes when the process runs as a General Input workflow?
The intake, the template, and the formatting rules become shared infrastructure instead of personal habits. Anyone forwards the email or drops the message in Slack and gets identical output. When the brand or boilerplate changes, you update the template once and everyone's next document is already right.
### Does the AI write or invent any of the content?
No. In the real estate example, every number on the sheet traces back to the MLS export and the photos come from the shared folder. The job is conversion, not generation, which is why the output is trustworthy enough to send without a rewrite pass.
---
# OAuth scopes are the wrong unit of control for AI agents
> OAuth scopes bundle a whole service into one consent switch. Hand that to an AI agent and it inherits everything. The fix is to control operations, not scopes.
Published: 2026-06-29. Last updated: 2026-06-29.
Connect Google Sheets to an agent and the consent screen gives you one real choice: "See, edit, create, and delete all your Google Sheets spreadsheets." All of them. There's a read-only version, and that is the entire menu ([Google for Developers](https://developers.google.com/workspace/sheets/api/scopes)).
That checkbox isn't a permission. It's a whole category of permissions, bundled and handed over as one. It's a key, not a job description. The key opens the door. It says nothing about what you're allowed to do once you're inside.
For about thirty years, that was fine. A person was always holding the key.
## A scope was sized for a human
You have the Sheets key, but you only ever touch the three cells you meant to. The scope is wide. Your behavior is narrow. Nobody lost sleep over the gap between the two, because you were the one minding it. The consent screen could afford to be lazy. You were the real access control.
This is why OAuth feels coarse the moment you look closely. It was designed for apps a person drives, where broad scopes plus human judgment add up to safe ([Strata](https://www.strata.io/blog/agentic-identity/why-agentic-ai-demands-more-from-oauth-6a/)). Take the person out and only the broad scope is left.
## An agent holds the whole keyring
Hand that key to an agent and the gap becomes the whole story. The agent inherits everything the scope allows. Ask it to append one row to one tab, and you have also authorized it to overwrite, reformat, and delete every spreadsheet in the account. It probably won't. "Probably" is carrying a lot of weight in that sentence.
This is the defining security problem of the agentic era, and the numbers have caught up to it. The State of AI Agent Security 2026 report found that 88% of organizations had a confirmed or suspected AI agent incident in the past year, and that nearly half of teams still wire agents up with shared API keys ([Gravitee](https://www.gravitee.io/blog/state-of-ai-agent-security-2026-report-when-adoption-outpaces-control)). OWASP added a principle called "least agency" to its 2026 Top 10 for Agentic Applications: give an agent the minimum autonomy its task needs, so a hijacked one can't do unlimited damage ([Aikido](https://www.aikido.dev/blog/owasp-top-10-agentic-applications)).
Least agency is the right idea. The problem is that the credential can't enforce it. Google does not mint a "this agent, write only, this one sheet" token. The scope is the smallest unit it sells.
## The permission you actually want is the operation
So move the decision down a level. Below the credential, to the operation. Not "can this reach Sheets" but "what is it allowed to do there." Read only. Write only. One endpoint, and nothing else.
Once the operation is the unit, the things you actually want become sentences you can write down:
- An agent that sends email on your behalf but can never open your inbox.
- One that reports on your CRM but can't change a single record.
- One that watches your Stripe revenue but can't move a dollar.
A scope can't express any of those. An operation-level permission expresses all three. That's what General Input Agent Firewall does. It filters every call at the level of the operation, below the credential: you choose what an agent can do independently of what the key can access, per connection when one agent needs a tighter leash than another, or across your whole workspace as a standing rule, spanning the 80,000+ tools in the catalog.
## Enforced below the key
Where you enforce a rule matters as much as the rule. General Input Agent Firewall runs at the runtime proxy that sits between the agent and the outside world, underneath the credential. A blocked call never leaves the sandbox, so it never reaches the model, so the model never gets the chance to be talked into making it.
And the key itself stays out of reach. The AI never sees it. Credentials are encrypted at rest, injected at runtime, and stripped before anything touches the model. Every call that does go out is written to an audit trail. The firewall decides which calls are even on the table.
Scopes answered the right question for the world that built them, the one where a human was always the last line of defense. Agents removed the human. The permission model has to follow the action down to where it actually happens.
Don't hand an agent a key. Hand it a list of operations.
## FAQs
### Why aren't OAuth scopes enough to secure an AI agent?
Scopes were designed for apps a human drives, so they bundle a whole service into one consent switch. "Connect Google Sheets" grants "see, edit, create, and delete all your spreadsheets." A person only ever touches the cells they meant to, so the gap between the broad scope and the narrow action never mattered. An agent inherits the whole scope, and the gap becomes the whole story.
### What is General Input Agent Firewall?
It moves the unit of control below the credential, down to the operation. You allow or block specific operations (read only, write only, a single endpoint) per connection or across your whole workspace, independent of what the key can technically reach, spanning 80,000+ tools.
### Where is the rule enforced?
At the runtime proxy that sits between the agent and the outside world, beneath the credential. A blocked call never leaves the sandbox, so it never reaches the model, so a prompt-injected agent cannot be talked into making it. The credential is injected at runtime and stripped before anything touches the model, so the AI never sees the key.
### Can I set different rules for different agents?
Yes. Configure it credential by credential when one agent needs a tighter leash than another, or set a standing rule across your entire workspace. Every call that does go out is written to an audit trail.
---
# Block Built an Operating System for Internal Tools. Every Company Will Run One.
> Block App Kit lets any employee build a tool from a sentence and ship it safely: an early look at the agentic operating system every company will run on.
Published: 2026-06-27. Last updated: 2026-06-27.
An operating system exists so you don't have to think about the hardware. You double-click a program. You don't allocate its memory, arbitrate its access to the disk, or check whether it's allowed to open the network. The kernel does all of that underneath, the same way, every time. The whole point is that running a program is safe and boring by default.
Block just shipped an operating system for internal tools, and the [writeup](https://engineering.block.xyz/blog/from-localhost-to-launched-safely-shipping-apps-that-anyone-can-build) reads exactly like one.
It's called Block App Kit. An employee describes a tool in plain language, an agent builds it, and it ships to coworkers in hours. Since launching in March, the catalog passed a thousand distinct apps, weekly active users climbed from the low hundreds into the thousands, and about 80% of the people building are in Sales, Support, Legal, and Finance -- not engineering. Jack Dorsey called it the [fastest-adopting tool](https://x.com/jack/status/2070799572687876100) in the company's history. The apps are the programs. The platform is the kernel.
## The point is who's building
Look again at that 80%. The people shipping these apps are in Sales, Support, and Finance, and that's the whole reason to build something like this. For years, "I need an internal dashboard" meant filing a ticket and waiting in the engineering queue, behind everything with a customer attached. The tool was three weeks out, if it came at all.
Block App Kit deletes the queue. Anyone describes the tool they need and has it that afternoon, and the leverage lands on everyone who used to depend on engineering for a five-line internal app. That's why a company bothers to build this: not to generate apps, but to let every person operate without an engineer in the loop.
## What the kernel actually does
The impressive part of Block App Kit isn't the part that writes the app. It's the part that lets a freshly generated app touch real systems without anyone getting hurt. Who's allowed to open it. Which services it can reach. Where the API keys live, and how an app uses one without ever holding it. What happened, recorded, on every call out. Authentication, authorization, secret management, data-access controls -- on by default. Block's own security org started pointing people to it as the sanctioned way to ship.
That's a kernel. It manages the dangerous shared resources -- credentials, data, the network -- so the thousand apps stacked on top don't each have to reinvent how to handle them safely. Strip that layer out and you don't have an operating system. You have a thousand programs free-handing access to production.
## Why every company ends up here
The demand for this is not a Block-sized phenomenon. Gartner counts [16.2 million citizen developers in 2026](https://searchlab.nl/en/statistics/no-code-low-code-statistics-2026), most of them outside IT, and projects [75% of new applications](https://kissflow.com/low-code/gartner-forecasts-on-low-code-development-market/) will be built with low-code by year's end. Every company already has people who could describe a useful tool in a sentence. What they don't have is the kernel underneath.
So the apps get built anyway, and they ship without one. A scan of 1,400 vibe-coded applications turned up [2,038 critical vulnerabilities, more than 400 leaked secrets, and 175 instances of exposed PII](https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-generated-code-vulnerability-surge-2026/), including medical and financial records. Across enterprises, AI-assisted developers ship code three to four times faster but introduce security findings at ten times the rate, and expose secrets at roughly double. That's what an OS-less world looks like. Fast, and leaking.
## Build the kernel or adopt one
Block could write its own. It has hundreds of engineers and can afford to pull a team off customer-facing work to build infrastructure that never ships to a customer. Ramp built its internal tooling in-house for the same reason -- the headcount was there, and the math worked. That's the rounding error of companies who can fund a kernel from scratch. For everyone else, a multi-quarter team dedicated to a tool no customer ever sees is exactly the project that doesn't get approved.
So they face the choice the industry has faced at every platform layer before this one. Nobody writes their own operating system anymore. Nobody writes their own database, or their own cloud. They adopt one and build on top, and the layer that used to be a moonshot becomes the floor everybody stands on.
## What the kernel looks like when you don't build it
General Input is that layer, available without the engineering project. The build step is the one you expect: someone describes the tool they need in plain language, and an agent assembles a working app from it. Everything wrapped around that build is the actual product.
Every app runs in an isolated sandbox, never on someone's laptop and never with a standing pipe into production. Share an app with someone in your workspace and it runs as you, using your connected accounts, so the person using it gets the result without ever touching a key or a token. The builder gets distribution, the user gets the tool, and the credential stays sealed off from both.
Underneath sits the same credential model Block had to engineer. The agent that writes the app never sees an API key. Credentials are encrypted at rest, injected at runtime inside the sandbox, and stripped before anything reaches the model. Below the credential, a proxy enforces what each connection is allowed to do, down to the individual operation across thousands of tools, so a call an app shouldn't make is blocked before it ever leaves the sandbox. Every call that does go out is written to an audit trail. Security isn't a setting someone remembers to switch on. It's the floor the apps run on.
That is the whole difference between a demo and a platform. A vibe-coded app impresses one person on one laptop. The same app on General Input can be handed to the entire company, touch the systems it needs to touch, and leave a record of everything it did, without a single credential ever passing through the model that built it.
## The operating system of the agentic company
Block proved the demand and the design in a single post. An employee describes a tool, an agent builds it, and it ships safely because the platform makes "safely" the default -- the way a good OS makes running a program something you stop thinking about.
These are the agentic operating systems of the future. Every company will run one. The app builder on top was the easy part. The kernel underneath was the engineering, and the only question left is whether you build it or it's already built for you.
---
# Q1 Belonged to Anthropic. The AI Lead Already Moved.
> In Q1 2026 Anthropic looked untouchable. One quarter later the frontier is a dead heat and OpenAI opened a new front below the model. The lead never sits still.
Published: 2026-06-24. Last updated: 2026-06-24.
Three months ago, the AI story had a clear protagonist. Anthropic had taken the enterprise. Claude Code held more than half the enterprise coding market against OpenAI's roughly one-fifth, and Anthropic had quietly passed OpenAI on both run-rate revenue and the share of businesses paying for it ([VentureBeat](https://venturebeat.com/technology/anthropic-finally-beat-openai-in-business-ai-adoption-but-3-big-threats-could-erase-its-lead)). If you were writing the narrative in March, it wrote itself: Anthropic had the upper hand, and it looked structural.
It wasn't. These leads almost never hold.
## The frontier became a dead heat
The first thing that happened is that the model gap closed. On the composite intelligence indexes, the top four models from Anthropic, OpenAI, Google, and xAI now sit within a handful of points of each other ([Artificial Analysis](https://artificialanalysis.ai/leaderboards/models)). Anthropic's newest Opus still edges the very top of one or two leaderboards, but "edges" is the point. A lead measured in fractions of a benchmark point is not a moat. It is a coin that lands heads this month and tails the next.
When every frontier model can read the contract, write the function, and run the agent loop about equally well, "we have the smartest model" stops being a reason to choose anyone. The labs know this. So they go looking for a different front to fight on.
## This morning, the front moved again
That is what OpenAI announced today. Not a smarter model. A chip.
OpenAI and Broadcom [unveiled "Jalapeño,"](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/) a custom processor built for one job: serving large language models faster and cheaper than the GPUs everyone rents. Broadcom's CEO pegged the savings at [roughly 50% versus typical AI GPUs](https://www.cnbc.com/2026/06/24/openai-and-broadcom-reveal-jalapeno-first-ai-chip-in-partnership.html), with engineering samples already running production workloads and deployment targeted for the end of the year.
Read it as a strategic move, not a hardware spec. OpenAI is no longer trying to win only on what the model knows. It is trying to win on what the model costs to run. That is a brand new axis, and Anthropic's Q1 coding lead does nothing to defend it.
## Why the cost front favors OpenAI
The shift below the model plays to OpenAI's strongest hand. Winning on inference cost is not a research problem, it is a capital and infrastructure problem, and almost nobody can move money and concrete as fast.
[Stargate](https://openai.com/index/announcing-the-stargate-project/), OpenAI's joint infrastructure venture, is a $500 billion program that has already crossed 10 gigawatts of planned capacity. A custom inference chip is exactly what you build when you have that much compute to fill and every point of efficiency compounds across all of it. Meanwhile the broader trend is doing OpenAI's argument for it: inference costs have been falling roughly [10x per year](https://a16z.com/llmflation-llm-inference-cost/), turning the model from a scarce resource into something closer to a utility. A chip that is half the cost is how you stay ahead of a curve that is already collapsing.
So expect OpenAI inference to get markedly faster and cheaper, and expect Anthropic, Google, and xAI to answer below the model too, with their own silicon and their own datacenter deals. The competition does not stop. It just relocates.
## The landscape is the product now
Here is the part that matters if you have to make decisions on top of all this. There is no durable frontier lead anymore. Q1 was Anthropic's, on quality. Q2 is being fought on cost. Whatever Q3 turns on, it will be something else again, and the name at the top will have changed at least once more before anyone agrees on why the last one was permanent.
Betting on a single lab is just betting that this quarter's leader stays the leader, and the last eighteen months say they won't. So assume the lead keeps moving. Build for a world where the model underneath you is swappable, because it is going to get swapped, on quality or on price, whether you plan for it or not.
Q1 belonged to Anthropic. Today OpenAI changed what the race is about. Whoever's winning right now, give it a quarter.
---
# Stop building one giant agent. Build a dozen small ones.
> GitHub's COO hasn't seen AI wipe out whole job functions, and neither have we. What works: small agents that each do one thing and stay easy to maintain.
Published: 2026-06-03. Last updated: 2026-06-03.
Everyone keeps talking about AI wiping out entire job functions. I still haven't seen that happen. Apparently, neither has GitHub's COO.
What does seem to work is much narrower. You take one repetitive task off your plate, train an agent to handle that one thing, and let it run. Then you move on to the next one. **None of these are especially impressive by themselves. But they add up.**
## The era of the perfect mega-skill is ending
For a while, every other post was "go download this skill," the perfectly managed thing that runs your entire workflow end to end. One massive, beautiful, all-knowing system that stitches a dozen tools together and produces some mega output. Whoever built the best one internally became the resident AI influencer.
That version mostly doesn't exist for us anymore. What we keep reaching for instead are incredibly micro agents that do one thing for us very, very well. Not "produce the full report." More like "identify the single most important piece of marketing information in this pile." One job, done cleanly.
Small and replaceable beats big and impressive. If you can't describe what an
agent does in one sentence, it's probably trying to do too much.
## A dozen small agents quietly run your week
The magic isn't in any single agent. It's in the stack of them. One drafts the follow-up. One files the receipts. One watches a channel and surfaces only the thing you'd actually want to see. Each one is boring on its own.
A dozen small agents later, a meaningful chunk of your week is happening without you touching it. That's the part nobody screenshots, because it isn't a flashy demo. It's just work that stopped landing on your desk.
This is exactly the shape General Input is built for. Each of those agents is one workflow you describe to a specialist in plain English, "every weekday at 7am, do this one thing." They sit side by side in a single workspace, each on its own schedule or trigger, and you add the next one whenever a new task starts eating your time. Nobody has to maintain a monolith, because there isn't one.
If you've never built one, [start with a single agent](/blog/how-to-build-your-first-agent), or grab one of the ready-made templates and point it at your tools. Resist the urge to make it do five things. The first running agent teaches you more than any course.
## The real reason small wins: you can fix it
Here's the part that matters most. The trick is keeping each agent small and easy to fix.
When the work changes, and it always changes, you want to tweak one piece. With a mega-skill, weeks go by, months go by, things shift, and now you're trying to surgically edit a giant all-knowing system you stopped understanding months ago. You're screwed. With a small agent, you open the one thing, change the one line, and move on.
On General Input that change is usually a sentence. A small agent is a short plain-English instruction, not a tangle of code, so adjusting it means telling the specialist what's different now, not refactoring a system. And because each one is cheap to build and cheaper to throw away, you're never precious about the agent that stopped being useful.
This is also why I'm not worried about AI erasing whole roles. The work that's actually getting automated is granular. Someone still has to decide which task to carve off next, what good output looks like, and where a human check belongs. That's [closer to operating than to being replaced](/blog/agent-operator-million-jobs).
So don't go hunting for the one perfect skill. Build the small one. Ship it. Then build the next.
## FAQs
### Why not just build one agent that does everything?
Because the work changes. A few months in, a mega-agent becomes a giant system you no longer fully understand, and one small change means untangling all of it. Small agents stay easy to read, easy to fix, and easy to throw away.
### How small should each agent be?
One job. It pulls the most important marketing signal, drafts one kind of reply, or files one kind of document. If you can describe what it does in a single sentence without an "and," it is the right size.
### Do a bunch of small agents actually add up to real time savings?
Yes, because they compound. None of them is impressive alone. But a dozen of them later, a meaningful chunk of your week is happening without you touching it.
---
# What is a local AI model? NVIDIA and Microsoft just put one in a laptop.
> NVIDIA and Microsoft announced laptops that run AI on the device. What a local model is in plain English, and why local-first and model-agnostic matter.
Published: 2026-06-01. Last updated: 2026-06-01.
Overnight at Computex in Taipei, during his GTC keynote, NVIDIA CEO Jensen Huang announced a partnership with Microsoft. They're launching a new line of Windows laptops built around a chip that runs powerful AI right on the machine, no connection to a far-off datacenter required ([NVIDIA](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-and-Microsoft-Reinvent-Windows-PCs-for-the-Age-of-Personal-AI/default.aspx), [Axios](https://www.axios.com/2026/06/01/microsoft-nvidia-surface-ultra-rtx-spark)).
If you're not deep in the AI world, that probably sounds like every other hardware announcement. But the phrase doing the heavy lifting is "local model," and it's worth understanding, because it changes where your data goes every time you use AI.
## How AI works for most people today
When you type a question into a popular AI tool today, your words leave your device and travel to a company's servers. The AI does its thinking there, on hardware you'll never see, and the answer travels back to you.
Two things follow from that. You're renting the AI, so the moment you stop paying, it's gone. And you're handing over your data to use it. Your draft email, your contract, your company's numbers, all of it makes a round trip to someone else's computer. For casual use, nobody minds. For anything sensitive, that round trip is exactly the part people are nervous about.
## What a local model actually means
A local model flips the setup. Instead of living in a datacenter, the AI lives on your own computer. Your documents, your emails, and your questions never leave the laptop, because the thinking happens right there on the machine in front of you.
The practical difference is simple. Your data never leaves the building. The same draft or question that used to make a round trip to a stranger's servers now stays on the device sitting on your desk. That's what NVIDIA and Microsoft just made possible for a normal laptop, with enough horsepower to run a serious model on-device.
## The cost angle nobody mentions
There's a money story here too. Companies today often pay for AI by the seat, and for heavy users running AI all day, subscriptions to providers like OpenAI or Anthropic can climb into the thousands per employee per month.
With a local model, the AI comes bundled with the PC you already bought. You pay for the laptop once instead of paying a per-person fee forever. For a business with hundreds of employees, that's a very different math problem.
## Why this is aimed squarely at Apple
NVIDIA isn't entering an empty room. Apple's Macs have quietly become the favorite machine for people who want to run AI fast, privately, and securely on their own hardware, thanks to strong chips and a lot of fast memory.
NVIDIA, with Microsoft and Windows behind it, is going straight at that audience ([CNBC](https://www.cnbc.com/2026/05/31/nvidias-new-chip-to-power-fresh-line-of-windows-laptops-by-dell-hp.html)). The new laptops ship this fall from familiar names like Dell, HP, Lenovo, ASUS, and Microsoft's own Surface line. The pitch is the same one Apple has been making. Your AI, on your machine, under your control.
## The other freedom is not being locked into one model
Privacy is the headline, but there's a second freedom hiding in this announcement, and it might matter more over time.
For the last few years, using AI has looked like the early days of computing, when an office shared a single machine that everyone lined up to use. You picked one company, learned its quirks, fed it your data, and built your habits around it. If a competitor shipped something better next month, switching meant starting over.
Putting a capable model on your own laptop is the same shift the personal computer made. The power moves from a shared system you borrow to one that's yours. And because these machines run open models, you get to be model-agnostic. You pick the best AI for the task in front of you, and swap it the day a better one shows up.
Local-first answers where your AI runs. Model-agnostic answers which one runs.
The two together are what put you back in control of both your data and your
options.
## Running it safely is still an orchestration problem
Here's the part the announcement glosses over. Running a model on your laptop solves where the AI lives. It doesn't, on its own, solve what the AI is allowed to do.
A real task isn't just answering a question. It's checking your calendar, pulling a file, sending an email, updating a record. The moment AI starts doing actual work, it needs access to your accounts, which means your passwords and keys are in play. Keeping the model on your laptop means nothing if it still has to be handed your passwords, or if there's no record of what it touched.
That's the part we built General Input around. The AI does the work across your tools, but it never sees your passwords. They're encrypted at rest, injected only when a step needs them, and stripped before anything reaches the model. Every action it takes on your behalf is logged. Pair that with model-agnostic routing and you can keep a sensitive step on the open-weight model running locally, send the rest to whatever cloud model fits, and still guarantee no model in the chain ever held a credential or moved data without a record.
Privacy isn't only where your data sits. It's what gets to touch it, and
whether you can see what happened. A faster chip fixes the first. The
orchestration layer fixes the rest.
## The takeaway
A local model is a real step forward for privacy, and these laptops will put it in a lot more hands. But the deeper lesson is about control. Don't bet your work on a single company's servers, and don't marry a single company's model. Own where your AI runs, and which one you run.
## FAQs
### What is a local AI model?
A local model runs on your own device instead of a company's servers. The AI does its thinking on the machine in front of you, so your files and questions never have to leave it.
### Does running a model locally make it private?
It is a real step forward, because your data never makes a round trip to someone else's datacenter. But privacy also depends on what the AI is allowed to touch and whether anyone can see what it did. A model that still holds your passwords or keeps no record of what it accessed is not private just because it runs locally.
### Does General Input support local models?
Yes. General Input is model-agnostic, so you can route each step of a workflow to the model that fits, a local open-weight model for sensitive work or a frontier cloud model for everything else, and swap either one whenever a better option ships.
### How does General Input keep my data safe when the AI does real work?
Credentials are encrypted at rest, injected at runtime only when a step needs them, and stripped before anything reaches the model. The AI never sees your API keys, and every piece of data it accesses from an external system is recorded in an audit trail.
---
# Show Me the Incentive, I'll Show You the Outcome
> Cerebras CEO Andrew Feldman says the biggest barrier to enterprise AI adoption isn't data. It's lawyers and security teams. Here's why he's right.
Published: 2026-05-26. Last updated: 2026-05-26.
Andrew Feldman, CEO of Cerebras, was asked on 20VC whether data quality is the biggest inhibitor to enterprise AI adoption. He didn't hesitate. No. **It's lawyers and security teams.**
His argument is simple. These teams have a brutal incentive structure. Nobody gets promoted for saying yes. Everybody gets blamed when something goes wrong. A year where nothing happened? That's the dream. So when a technology shows up with zero precedent, zero case law, zero regulatory playbook, the rational move is to say no.
And he's right that lawyers are trained on precedent. Ask one to work in an area where nobody has done it before and they genuinely don't know what to do. Their whole toolkit is "what has everyone else done?" AI doesn't have a "what everyone else has done" yet.
## The data backs up the incentive problem
The data tells the same story. [Gallagher surveyed](https://www.ajg.com/news-and-insights/features/ai-adoption-and-risk-benchmarking-2026/) 1,200+ businesses and found **56% cite legal and reputational risk** as their top AI concern. [Gartner found](https://www.gartner.com/en/newsroom/press-releases/2025-10-06-gartner-predicts-ai-regulatory-violations-will-result-in-a-30-percent-increase-in-legal-disputes-for-tech-companies-by-2028) 70%+ of IT leaders rank compliance as a top-three blocker. [Deloitte](https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html) says only a third of orgs have anything resembling mature AI governance. Everyone else is still arguing about who gets to decide.
## Even NVIDIA had to force the issue
Feldman also mentioned Jensen Huang having to [battle his own internal teams](https://dataconomy.com/2025/10/15/jensen-huang-says-every-nvidia-engineer-now-codes-with-cursor/) at NVIDIA to get Cursor adopted. He eventually just decreed it. The CEO of the most AI-forward company on the planet had to force the issue.
That's the part that sticks with me. **If NVIDIA has this problem, everyone has this problem.**
The cautious approach has real costs too. Every month spent in a review cycle,
a competitor without a compliance committee is shipping, building
institutional knowledge, and compounding the advantage.
## The cost of waiting for consensus
But here's the thing I keep turning over. Feldman frames legal and security as a drag on adoption, and the incentive analysis is hard to argue with. But there's a real cost on the other side too. Every month you spend in a review cycle, a competitor without a compliance committee is shipping. They're building institutional knowledge, tightening feedback loops, compounding the advantage. That gap doesn't close easily.
So someone has to push the boulder. Not by going around legal or pretending the risks aren't real. But by **creating the precedent that legal needs to get comfortable.** Ship one thing. Make it work. Don't cause an incident. That's the playbook.
## Who actually pushes AI adoption forward
The tricky part is that most companies aren't set up for this. There's no role that owns "push AI adoption past internal resistance." It just kind of falls to whoever cares enough to fight for it. And fighting your own security team is exhausting, even when you're right.
I don't think there's a clean answer here. The cautious approach has real costs. The aggressive approach has real risks. But I think the companies that figure out how to move deliberately, not recklessly but not slowly either, are the ones that end up in a very different position two years from now.
**Waiting feels safe. But it's not free.**
---
# How to Build Your First Agent: A Five-Step Field Guide
> Five decisions that turn a vague AI idea into a working agent in production. No code required, just a clear trigger, a goal, and a feedback loop.
Published: 2026-05-21. Last updated: 2026-05-21.
We've been asking companies a simple question: **how many of your employees have actually deployed an agent** to automate part of their day? The answers tell the story.
"You mean chatting with Claude?"
"I have no idea what an agent is."
"The engineers love agents!"
That gap is the real story of AI adoption right now. [McKinsey's November 2025 State of AI](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) puts 23% of organizations scaling agentic AI and another 39% experimenting. Those numbers describe the company, not the people in it. At the employee level, the share who've shipped their own agent is a rounding error.
It's not because the technology isn't ready. The frontier models can already do the work. The reason most employees haven't deployed an agent is that **nobody has handed them a repeatable mental model** for what "deploying one" even looks like. So most attempts fizzle. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) that over 40% of agentic AI projects will be canceled by the end of 2027.
Agents are supposed to be a natural extension of your team, accomplishing tasks that would otherwise require a human. The formula for getting one into production is pretty simple. Five decisions, in order.
## 1. Pick the right trigger for your agent
The first decision is the most important and the most underrated. An agent is not a chat window. It's a process that runs when something happens. A new meeting note appears in Granola. A row lands in a spreadsheet. A webhook fires. A cron job kicks off every morning at 8am.
**Picking the trigger pins down what the agent is actually for.** "Summarize emails" is not a workable spec. "When a new Granola note is created, log the call in HubSpot and draft a follow-up email in Gmail" is one. Vague triggers produce vague agents.
## 2. Define a clear goal with guardrails
Once the trigger is set, write down what the agent should accomplish when it fires. Two or three sentences for the end goal, plus a few rules of thumb for how to get there. "Always reply in the customer's language." "Never quote a price." "If the request mentions a refund, escalate."
Write the instructions you'd give a smart new hire on day one. If you have to
explain every edge case in advance, you're building a flowchart, not an agent.
## 3. Pick the model and add human-in-the-loop
Different jobs need different lanes. A summarization agent doesn't need the same reasoning model as one negotiating a refund. A high-volume tagging agent doesn't need the same context window as one reading a 50-page contract. The model layer is changing fast enough that the right choice in March can be the wrong choice in May. Build agents on a platform that lets you swap.
In the same step, decide where a human belongs. The default for any agent touching real systems should be: **a person approves before anything irreversible happens.** Sending an email. Updating a CRM record. Charging a customer. Deloitte's research shows teams that [redesigned workflows around human-AI collaboration](https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html) saw a 30% productivity lift, compared to just 5% when AI was bolted onto unchanged processes. Human-in-the-loop is not a tax. It's the thing that lets the agent earn trust fast enough to eventually remove the loop.
## 4. Build the simplest first draft
With the trigger, goal, model, and approval points decided, build the agent. The first version should be **the simplest thing that could possibly work.** Resist the temptation to wire in five tools, three retrieval systems, and a self-critique loop on day one. You don't yet know what the agent gets wrong. Get something running end-to-end before you optimize anything.
This is also the right time to think about credentials. API keys should be
encrypted at rest, injected at runtime, and never visible to the model itself.
If your platform lets the AI see raw API keys, the blast radius of a single
bad agent is your entire stack.
## 5. Test on real inputs until the runs are boring
This is where most teams quit. The agent runs on three test cases, looks impressive, and gets called done. Production inputs are messier. The agent encounters a customer writing in two languages mid-thread. It hits a malformed webhook payload. It tries to escalate to a Slack channel that was archived last quarter.
Run the agent on real inputs. Watch it fail. Tighten the goal or the rules of thumb. Add a tool when a missing capability shows up twice. Move a step behind human approval when the agent makes a class of mistake you can't accept. **Keep going until the runs are boring.** Boring is the signal you're done.
## What "deployed" actually means
The first agent doesn't need to be complex. It needs to be useful. An agent worth running in production is one you've watched fail enough times to trust.
That's how agent adoption starts. Not with a company-wide transformation deck, but with one person deploying one useful agent into one real workflow.
## FAQs
### Do I Need to Know How to Code to Build an AI Agent?
No. You describe the agent in plain English. The platform handles the orchestration, integrations, and execution. The skill that matters is deciding which workflow is worth automating first.
### What's the Smallest Useful First Agent?
Pick the most repetitive part of your week with a clear trigger and a clear output. A new Granola note turning into a HubSpot log and a Gmail draft is a good shape. One real input, one real result, end-to-end.
### How Does Human-in-the-Loop Actually Work in a Workflow?
You decide which steps the agent runs unsupervised and where it pauses for approval. Anything irreversible -- sending an email, updating a CRM record, charging a customer -- should pause for a human until you trust the agent. Once it's been right enough times, you can drop the checkpoint.
### Will the AI Model See My API Keys?
No. Credentials are encrypted at rest, injected into the sandboxed step at runtime, and stripped from any payload before it reaches the model. The agent can call your tools without the model ever seeing the secrets.
---
# Cerebras hit 981 tokens per second on Kimi K2.6. The inference layer is now its own category.
> Cerebras served Kimi K2.6 at 981 tok/s, 29x faster than the official endpoint. Why inference speed changes everything for generative UI and AI workflows.
Published: 2026-05-21. Last updated: 2026-05-21.
On May 6, 2026, [Artificial Analysis measured Cerebras serving Kimi K2.6 at 981 output tokens per second](https://artificialanalysis.ai/models/kimi-k2-6/providers). K2.6 is [Moonshot AI's trillion-parameter open-weight model](https://www.marktechpost.com/2026/04/20/moonshot-ai-releases-kimi-k2-6-with-long-horizon-coding-agent-swarm-scaling-to-300-sub-agents-and-4000-coordinated-steps/), released April 20 with native multimodal handling and an agent swarm that scales to 300 sub-agents over 4,000 coordinated steps. The official Kimi endpoint serves the same weights, and the same 10,000-token request takes it 163.7 seconds. **Cerebras returns the answer in 5.6.** That is a 29x improvement in time to final answer on identical model weights.
The model is one decision. The inference provider is a separate one. The same
open weights can run on completely different infrastructure with wildly
different speed and cost profiles.
## What 981 tokens per second actually unlocks
For most of the last two years, **generative UI** has been a research demo with an asterisk. The idea, [outlined by Google Research](https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/), is that the model writes the interface itself for each prompt instead of routing the user through a fixed app. The asterisk has always been latency. A custom UI that takes a minute to render is a slideshow. A custom UI that streams in inside five seconds is software.
**AI-generated apps** work on the same arithmetic. A few thousand lines of code at a thousand tokens per second is a coffee sip. A few thousand lines of code at 30 tokens per second is a meeting. The product category doesn't change because the model got smarter. It changes because the inference got fast enough that a person stays in flow.
The Cerebras run is the first time those numbers have been hit on a frontier-grade open-weight model in production.
## How the inference layer split from the model layer
A year ago, choosing a model meant choosing an endpoint. The model lab built the weights, hosted the API, set the price, and that was the package. That has come apart. The same Kimi K2.6 weights now run on the Moonshot endpoint, [on Cerebras wafer-scale inference](https://www.cerebras.ai/blog/cerebras-kimi-k2-Enterprise), on NVIDIA NIM, on Azure AI Foundry, on GPU clouds like DeepInfra and Spheron, and on local hardware. Each one has a different latency, a different cost per token, and a different ceiling on context length.
This is the **second major axis of model choice** that opened up in 2026. The first was which weights to use, where DeepSeek V4, Kimi K2.6, GPT-5.5, and Opus 4.7 are now all live options depending on task. The second is which inference provider serves those weights, where Cerebras is [6.7x faster than the next-fastest GPU cloud and 23x faster than the median](https://artificialanalysis.ai/models/kimi-k2-6/providers) on this particular workload.
Both axes move on their own timeline. A new frontier model lands every few weeks. A new inference provider posts a new throughput record every few months. The platform that locked in last quarter is already a generation behind.
## Why pinning to one inference endpoint is a planning error
Most workflow tools still bind to a single provider per workflow. That made sense when the model and the endpoint were the same decision. It doesn't survive the split.
A real workflow has steps with different shapes. A planning step wants the smartest reasoning model and is fine waiting 30 seconds. A UI generation step needs 1,000 tokens per second or the interaction breaks. A summarization step running over 50,000 documents wants the cheapest per-token rate that clears a quality bar. Routing all three through the same endpoint means overpaying on the bulk step, waiting too long on the UI step, or losing quality on the planning step. Usually all three.
Per-step routing solves this. Each step in a workflow declares the constraint
that matters, and the platform picks the model and provider that satisfy it.
The fix is **per-step routing**. Each step in the workflow declares the constraint that matters for that step. The platform picks the model and the provider that satisfy it. When Cerebras lands a faster Kimi, the UI step picks it up without anyone rewriting the workflow. When DeepSeek's next release undercuts the bulk step on cost, the bulk step picks it up. The workflow author writes the logic once and the substrate keeps getting faster and cheaper underneath them.
## What this means for the next twelve months
The next year of model innovation is going to look like the last six weeks. A new frontier model approximately every release cycle. A new inference benchmark approximately every month. Cerebras itself just [IPO'd at a $95B market cap on May 14](https://www.theregister.com/ai-ml/2026/05/15/cerebras-wafer-scale-ai-bet-delivers-blockbuster-ipo/5240821) on the bet that this trajectory continues. There are too many independent axes of improvement to predict which combination wins.
The teams that compound the fastest are the ones whose workflows can adopt each release without a rewrite. **The model isn't the bottleneck anymore. The inference provider is.** And both of them are moving too fast to bet on.
## FAQs
### How fast is 981 tokens per second compared to a typical model endpoint?
It's roughly 29x faster than the official Kimi endpoint serving the same K2.6 weights, and about 23x faster than the median GPU cloud on that workload. A 10,000-token response that takes 163 seconds on a standard endpoint returns in under 6 seconds on Cerebras.
### Does picking a faster inference provider mean using a different model?
No. The same open-weight model can run on multiple providers, each with different latency, cost, and context-length tradeoffs. The model is one decision, the inference provider is a separate one.
### Why does inference speed matter for generative UI and AI-written apps?
Custom interfaces and generated code are only usable if they render fast enough to keep a person in flow. Sub-10-second response times turn slideshow demos into software people actually use.
### Can I route different steps in a single workflow to different models and providers?
Yes. General Input lets each step in a workflow pick the model and inference provider that fits its constraints, so a planning step, a UI step, and a bulk summarization step can each use the substrate that's right for them.
---
# If your best automation lives on one laptop, you don't have automation
> Share AI workflows and credentials across your team. Move automations off individual machines and into the org so they survive turnover.
Published: 2026-05-07. Last updated: 2026-05-07.
Every team has one person who has figured out the automation. Their workflow is good. It saves real hours. It's wired up to their personal credentials, it runs from their account, and it lives in their head. Nobody else on the team uses it. Sometimes nobody else even knows it exists. When they change roles or leave, the playbook walks out with them.
This is the **tribal-knowledge problem, applied to automation**. A recent analysis found that [42% of institutional knowledge resides solely with individual employees](https://www.iteratorshq.com/blog/cost-of-organizational-knowledge-loss-and-countermeasures/) and that [more than 70% of process knowledge still lives in people's heads, emails, or spreadsheets](https://blog.rapidautomation.ai/how-manufacturers-can-solve-tribal-knowledge-bottlenecks-and-improve-operational-consistency/). IDC pegs the cost of poor knowledge sharing at $31.5B per year. The personal-laptop automation is a small, daily version of the same failure.
Today General Input ships **native sharing for workflows and credentials**. This is what fixes it.
## Why most workflow automations stay personal
Most automation tools were built around the individual builder. Each person sets up their own connections, builds their own workflows, troubleshoots their own runs. Sharing is usually a settings page nobody opens, or it's gated behind an enterprise tier. Best practices don't propagate, because the tool itself isn't shaped around a team. It's shaped around an account.
So the muscle memory becomes: build it on my machine, run it from my account, never hand it off. Even when somebody asks for help, the path of least resistance is to DM them an API key or screen-record the build. That's not knowledge transfer. That's a copy.
## How workflow and credential sharing works
When you share a workflow with someone in your workspace or your whole organization, the credentials that workflow depends on travel with it. The person you share with can run the automation, but they **never see the underlying tokens**. The credential stays owned by you.
At runtime, the executor's identity is what shows up in the audit trail. Their account is what triggered the run. But the OAuth token, the API key, the database password are injected into a sandboxed step at execution time and stripped before any payload reaches the AI. The model performing the work never sees a secret.
Practically, this means one person sets up the Salesforce connection once, builds a workflow on top of it, and the rest of the team uses it like an internal tool. If the person who built it leaves, the workflow keeps running. **The automation belongs to the org, not to a laptop.**
One person sets up the connection and builds the workflow. Everyone else on
the team runs it like an internal tool, without needing their own admin
credentials.
## How sharing differs from templates
General Input has had templates since launch. Templates are the **bring-your-own-credential pattern**: a workflow shape gets published, and each person who instantiates it plugs in their own connections. Useful when the automation is generic.
Sharing solves a different problem. Sharing is "use my setup." When the workflow is calling your CRM, with your permissions, on a schedule you defined, you don't want everyone on your team to also need their own admin credentials to your CRM. You want them to be able to run your automation. That's what sharing now does.
## The credential visibility invariant
There is one architectural rule worth calling out. **A workflow's visibility cannot exceed the visibility of any credential it references.** If a workflow uses a credential that's only shared with three people, the workflow itself can't be shared with the whole org. Sharing the workflow more broadly forces the underlying credential to be shared more broadly first.
This is the part most "team plan" features get wrong. They treat sharing as a
permissions toggle on the workflow and let the credentials lag behind, which
leads to broken executions for half the team. General Input models it the
other way: the credential's visibility is the floor.
## The real shift is institutional memory
The framing isn't really security, though that's there. The framing is institutional memory. The best automation on the team should not be one person's setup. It should be something the whole org runs, something that survives turnover, something a new hire can use on their first day.
If your best automation lives on one laptop, you don't have automation. You have a person.
Sharing is live on every plan today.
## FAQs
### Does the person I share with see my underlying API keys or OAuth tokens?
No. Credentials stay encrypted at rest and are injected into the sandboxed step at runtime under your ownership. The person you share with can run the workflow but never sees the secret.
### What happens to a shared workflow if I leave the company?
The workflow keeps running. It belongs to the org, not your laptop, so audit trails, executions, and ownership transfer cleanly without rebuilding the automation.
### How is sharing different from publishing a template?
Templates are bring-your-own-credential, meaning every user plugs in their own connections. Sharing is "use my setup": your credentials power the workflow for everyone you share it with.
### Can I share a workflow with the whole org if its credential is only shared with three people?
No. A workflow's visibility is capped by the visibility of every credential it references. Expanding the workflow's reach forces you to expand the underlying credential's reach first.
---
# General Input joins the Microsoft AI Cloud Partner Program
> General Input is a certified Microsoft AI Cloud Partner. Connect Outlook, Teams, OneDrive, SharePoint, Excel, Planner, and To Do to AI workflows in one click.
Published: 2026-05-04. Last updated: 2026-05-07.
General Input is now a member of the [Microsoft AI Cloud Partner Program](https://learn.microsoft.com/en-us/partner-center/membership/isv-success). With this release, **seven Microsoft 365 services are 1-click connect** on any General Input workflow.
## Supported Microsoft 365 services
- **Outlook** -- read, send, search email, manage calendar events
- **Teams** -- post messages, read channels, send DMs, manage meetings
- **OneDrive** -- list, read, write, and search files
- **SharePoint** -- access sites, lists, and document libraries
- **Excel** -- read and write rows in workbook tables
- **Planner** -- create, update, and read tasks across plans
- **To Do** -- manage personal task lists
Each is a **1-click OAuth connect** from the connector dropdown on any workflow, with permissions scoped per service.
Your IT team does not need to register an app inside your tenant. The
Microsoft AI Cloud Partner certification means General Input is the registered
app. Users authorize via Microsoft's standard consent screen.
## How Microsoft credentials are secured
The partner review covered how General Input handles credentials, and it's worth describing briefly because the architecture answers a question every agent buyer is starting to ask.
OAuth tokens are **encrypted at rest** in our credential store. When a workflow runs, the credential is decrypted and injected into a sandboxed step at runtime, then destroyed when the step finishes. The token is exposed to the Microsoft Graph SDK call and nothing else. It is stripped from any payload before it reaches the model. General Input can now perform real work across your Microsoft org, and the model driving any of it never sees a credential.
Every Graph call is written to an audit trail tied to the user who authorized
the credential, the workflow that ran, and the step that made the call.
## Available today
All seven Microsoft connectors are live for every General Input workspace.
## FAQs
### Which Microsoft 365 services are supported with one-click OAuth?
Outlook, Teams, OneDrive, SharePoint, Excel, Planner, and To Do. Each is a per-service connector with permissions scoped to that surface.
### Does the AI model see my Microsoft access tokens?
No. Tokens are encrypted at rest, decrypted only inside a sandboxed step at runtime, and stripped from any payload before it reaches the model.
### Will my IT or InfoSec team have to register an app inside our tenant?
No. The Microsoft AI Cloud Partner certification means General Input is the registered app, so your users authorize via Microsoft's standard consent screen.
### Is every Graph call audit-logged?
Yes. Each call is tied to the user who authorized the credential, the workflow that ran it, and the step that issued it.
---
# Don't build your workflows inside a single lab's application
> Model-agnostic workflow automation lets you swap AI models per task with zero rewrite. Stop locking workflows to one lab's pricing and roadmap.
Published: 2026-04-24. Last updated: 2026-05-07.
Imagine you signed a 24-month enterprise AI agreement on Monday, locked to one model lab for all your AI implementation needs.
Then yesterday happens. [GPT-5.5 drops](https://www.buildfastwithai.com/blogs/gpt-5-5-review-2026) claiming to beat Opus 4.7 on reasoning and leading Terminal-Bench 2.0. Today, [DeepSeek V4 Pro drops](https://simonwillison.net/2026/Apr/24/deepseek-v4/) at [roughly one-tenth the price](https://venturebeat.com/technology/deepseek-v4-arrives-with-near-state-of-the-art-intelligence-at-1-6th-the-cost-of-opus-4-7-gpt-5-5) of either when you factor in cached inputs, within a few benchmark points on SWE-bench Pro and ahead on sub-tests like LiveCodeBench and Codeforces.
**Workflows built inside a model provider's application are locked to their roadmap and their pricing.** Not the priorities of your business.
## What two frontier model launches in 48 hours revealed
GPT-5.5 leads Terminal-Bench 2.0 at 82.7 versus V4 Pro's 67.9 and claims the top reasoning score over Opus 4.7. V4 Pro Max leads LiveCodeBench and Codeforces. Opus 4.7 still tops SWE-bench Pro at 64.3. V4 Pro output runs about $3.48 per million tokens against Opus 4.7 at $25 and GPT-5.5 at $30, and the gap widens to roughly 1/8th and 1/10th with cached input.
None of these three models is the right answer for every task. That is exactly
why a 24-month single-lab contract is a bet you don't need to make.
## How Cursor taught engineers to stop betting on one AI lab
Engineering teams already solved this. They stopped betting on a lab and started betting on the editor that sits between all the labs.
Cursor is that editor. Every frontier model is a dropdown. **Switching cost per task is zero.** When a new model ships the engineer tries it the day it lands. If it's better, they move. If it's worse, they don't. Cursor is [now used by 70% of the Fortune 1000](https://tech-insider.org/cursor-60-billion-valuation-anysphere-ai-coding-2026/). The labs compete for time inside it, which keeps every one of them honest on price and quality.
Workflow automation needs the same layer. Not a different flowchart tool. A different relationship between the workflow definition and the model that runs it.
## The rewrite tax teams pay when locked to one AI provider
If your automations call one lab's API directly, or sit on an orchestration tool that hardcodes a single provider, you cannot capture the gains from either of this week's releases without a rewrite. [A recent Zapier survey](https://zapier.com/blog/ai-vendor-lock-in-survey/) found **58% of enterprise AI vendor migrations either failed or took far more effort than expected**, and nearly three-quarters of enterprise leaders said losing their current AI vendor would disrupt day-to-day operations.
So most teams skip the migration. They keep running last year's model choice on this year's workload. The tax is worst on the workflows you run most. A 10,000-row enrichment job that ran fine on Opus last quarter is now paying seven to ten times the cost it needs to for output a cheaper model would produce. A daily pipeline digest that genuinely benefits from Opus's reasoning is still worth the spend. The problem is that a single-lab contract gives you no way to treat those two workflows differently.
A 10,000-row enrichment on DeepSeek V4 Pro costs a few dollars. The same job
on Opus 4.7 costs around forty. Model portability is not abstract. It is a
line item.
## What model-portable workflows look like in practice
On General Input every workflow is **model-portable by default**. You describe the workflow in plain English. At run time you pick the model. Per template, per run, per tier inside a single run. Credentials for each lab are stored once, encrypted at rest, injected at call time, and stripped before any output reaches the model that's generating text. The workflow file is the same whether it calls Opus 4.7, GPT-5.5, or DeepSeek V4 Pro.
That means the 10,000-row enrichment job can run on DeepSeek V4 Pro for a few dollars instead of forty. The daily pipeline digest can stay on Opus because the reasoning matters and it runs once. The ambiguous-ticket escalator can classify cheaply and route the hard cases to whichever model you trust most this week. When the next frontier model drops next month, you add it to the list and keep going.
## The labs will keep leapfrogging
They always do. The release cycle is shorter than the useful life of any recurring workflow, and that will stay true. Spend your energy finding the right workflow for your business. The model underneath will keep getting faster and cheaper without you.
DeepSeek V4 and V4 Pro are live in General Input today: pick either from the model dropdown on any workflow.
## FAQs
### Can I switch the model on a workflow without rewriting it?
Yes. Workflows are described in plain English; the model is a dropdown you can change per template, per run, or per tier inside a single run.
### How does General Input handle credentials for multiple labs?
Credentials for each provider are stored once, encrypted at rest, injected at call time, and stripped before any output reaches the model that's generating text.
### Which models are available today?
Frontier models from every major lab, including Claude Opus 4.7, GPT-5.5, Gemini, and DeepSeek V4 / V4 Pro. Plus any OpenAI-compatible endpoint via BYO model deployment.
### Can I route cheap tasks to one model and reasoning-heavy work to another?
Yes. You can pick a different model per workflow, or even per step inside a workflow, so a 10,000-row enrichment can run on a cheap model while a daily briefing stays on a frontier reasoning model.
---
# The agent fabric has to reach the systems most of the economy actually runs on
> Most agent platforms stop at Google, Notion, and HubSpot. We shipped Bullhorn and LeadMagic because agents need to reach the long tail of business software.
Published: 2026-04-23. Last updated: 2026-05-07.
Most agent-builder platforms seem to assume every business runs on Google, Notion, and HubSpot.
**Most of the economy runs on something else.**
Two of our clients asked us to integrate Bullhorn Recruiting and LeadMagic. Both were paying for Claude and ChatGPT subscriptions that had no native way to reach either one. The AI was there. The data was there. Nothing connected them.
Today we shipped both.
## Why Agent Platforms Leave the Long Tail Behind
If you work at a tech company, the last two years have been an unbroken wave of automation options. Chain tools together in Zapier. Call the OpenAI API from a Python script. Point Claude at MCP servers that map to your stack. Buy an agent platform with pre-built connectors to every SaaS tool on your desk. The examples online are always the same: Google Calendar, Notion databases, HubSpot deals, Slack threads.
The rest of the economy has been watching from outside. **Staffing runs on Bullhorn. Real estate runs on Follow Up Boss.** Legal firms run on their own matter-management software. Construction runs on scheduling systems you have never heard of. Healthcare operations, logistics, field service: each one has a CRM that is the center of the workday and nowhere on the big model providers' radar.
Buying a frontier-model seat does not close that gap on its own. The model is capable. The problem is reach. Neither Claude nor ChatGPT has a native path into Bullhorn or LeadMagic. A recruiter paying for an Enterprise seat every month can ask the model anything except the thing they actually need answered: what is going on in the system where my work lives.
A recruiter paying $20/month for an AI subscription still can't ask it the one
question that matters: what's happening in Bullhorn right now?
## How Two Clients Used Bullhorn and LeadMagic Together
The first client runs a staffing firm and wanted to **reactivate dormant candidates**. Their Bullhorn database has over ten thousand candidates they have talked to at some point. Most of them sit there untouched. [Bullhorn's 2026 GRID report](https://www.bullhorn.com/grid/2026-industry-trends/report/) says 70% of contractors are never redeployed after their last assignment. Reactivation is not one task. It is five: segment dormant candidates, find a current mobile number, draft a personalized message, send it through a channel that actually gets a response, log the outreach back so the record stays clean. [LeadMagic](https://leadmagic.io) handles the mobile-number step better than the general-purpose enrichment providers. Bullhorn is the system of record. The AI does the drafting and the routing. None of that works unless all three are connected.
The second client runs outbound sales inside an agency and needed LeadMagic for the same reason every outbound seller needs it: **verified mobile numbers on decision-makers** with much higher match rates than the legacy data providers. They were building prospect lists in a CRM the frontier models cannot see, dumping them into a spreadsheet, and pasting enriched numbers back in by hand.
Both of them were paying full price for AI subscriptions that could not help.
The AI was capable. The data existed. Nothing connected them. That's the
problem one-click integrations solve.
## Building the Agent Fabric Industry by Industry
General Input has been working through a backlog of integrations industry by industry. Bullhorn and LeadMagic are this week. Follow Up Boss for real estate shipped a while back. Attio, Airtable, and the Microsoft stack are already live. Staffing, recruiting, real estate, legal, field service, healthcare operations: every sector has a system of record that belongs on the agent fabric. We are shipping toward that, not toward a single vertical.
If there is a system your business runs on that still doesn't talk to AI, the fastest way to get it supported is to tell us. We have been shipping client by client, which is slower than promising ten integrations a week on a homepage and faster than almost every other platform actually delivers. The homepage carries the current list, and it gets longer every week.
**The agent fabric can't be for tech companies only.** If a business has a system of record and a workflow worth running, agents should work there. That's what we are building toward.
## FAQs
### My company runs on a niche system of record. Can I still automate with General Input?
Probably yes. We've been shipping integrations industry by industry, including staffing (Bullhorn), real estate (Follow Up Boss), and outbound (LeadMagic). If yours isn't live yet, tell us and we'll prioritize it.
### How do new integrations usually get added?
Client by client, prioritized by what teams actually need. The current list is on the homepage and updates weekly.
### Do these long-tail integrations get the same security guarantees as Google or Microsoft?
Yes. Every integration uses the same credential-handling architecture: encrypted at rest, injected at runtime, stripped before the model sees anything.
### How do I request a new integration?
Reach out to us with the system you need. We've shipped requested integrations within days when the API is open.
---
# The Agent Operator: The Million-Job Role Aaron Levie Sees Coming
> Box CEO Aaron Levie predicts agent operator will be a million-job role. Here's what the job looks like inside regulated enterprises today.
Published: 2026-04-21. Last updated: 2026-05-07.
Everyone is afraid of AI making jobs disappear. Box founder Aaron Levie has a different take. In a clip this week he said **the biggest new job title of the next five years doesn't exist today**, and it's going to create a million roles. He calls it the agent operator.
Not the person using an agent. The person who walks into a marketing, legal, operations, or life sciences team and **redesigns the workflow so an agent can run most of it**. They're the interpreter between how work has been done and how it will be done.
The claim is easy to dismiss as another AI-future platitude. But his framing lands harder if you've actually tried to roll out agents inside a Fortune 1000 pharma, a bank, or a consultancy. Regulated. Data fragmented across fifteen systems. Employees wired to run a process a particular way for a decade. An agent can't inherit that. It has to be rebuilt for how an agent actually reads and acts.
## The Job Is Business Process Translation, Not Prompting
Most of the AI conversation is about the model. Which frontier model reasons best, which handles long context, which generates better code. But if you've tried to roll an agent into a function at a regulated enterprise, the model is almost never the bottleneck. **The bottleneck is that an agent can't inherit a process written for humans.**
A legal analyst flags a contract for reasons that aren't in the playbook. They're in her head from twelve years of reviewing deals. A marketer's attribution model is technically "last touch" but she knows to discount three specific channels because of a known spam problem nobody documented. A finance manager sanity-checks a number against a second source because a bad quarter in 2022 taught her the first source lies when volume spikes.
None of that is in the SOP. It's institutional knowledge. Someone has to translate it into explicit instructions, guardrails, data sources, and checks an agent can run. That translation is the job Levie is describing when he says "requires care and feeding and a real level of technical and business process acumen." It's not prompt engineering. It's business process redesign, with a model in the loop.
The hard part isn't the AI model. It's extracting the tribal knowledge that
lives in people's heads and encoding it into instructions an agent can follow.
## Why Agent Operator Will Be a Million-Job Role
Every function in every large company has to be redone. Marketing needs campaign performance workflows. Legal needs contract triage. Operations needs postmortem drafters. Research needs literature monitors. None of these are generic. Each one requires someone who understands the function, the data, and the agent. Multiply by every function at every large company and you get Levie's million.
[79% of enterprises face AI adoption challenges](https://writer.com/blog/enterprise-ai-adoption-2026/) despite massive investment. [Two-thirds of enterprise AI projects stall in pilot](https://hbr.org/2026/02/why-ai-adoption-stalls-according-to-industry-data). The model is good enough. **The people who can do the translation work are the bottleneck.**
The model is good enough. The bottleneck is the people who can translate
business processes into agent-ready workflows.
## The Agent Operator Role Already Exists (Without the Title)
The title isn't on the org chart yet. The work is. People inside large companies are already redesigning workflows for agents, writing instructions, wiring up integrations, and proving to legal that nothing sensitive ever reaches the model. They're doing it with a mix of Python scripts, Zapier, and internal platforms stitched together by IT. In five years they'll have a title and a team. For now they're doing the job without the name.
## FAQs
### What does an agent operator actually do day-to-day?
They redesign a function's workflow so an agent can run most of it. The work is translating tribal knowledge into explicit instructions, data sources, guardrails, and checks an agent can execute.
### Why isn't this just prompt engineering?
Prompt engineering tunes a model's response to a question. An agent operator redesigns the underlying business process, wires up integrations, and proves to legal and compliance that nothing sensitive reaches the model.
### Do I need engineering skills to be an agent operator?
No. The job is closer to business process redesign than software development. A platform like General Input handles the orchestration; the operator focuses on what the agent should do, what data it needs, and where the human checks belong.
### Where do I start if I want to be the agent operator inside my function?
Pick one workflow you currently run by hand, document the unwritten rules you apply to it, and rebuild it as an automation. The first running workflow teaches you more than any course.
---
# One-Click Integrations for the Whole Stack: Microsoft, Google, HubSpot, Salesforce, Reddit, and More
> General Input shipped one-click OAuth integrations for Microsoft, Google, HubSpot, Salesforce, and dozens more. Connect your tools in minutes, not hours.
Published: 2026-04-15. Last updated: 2026-05-07.
The biggest piece of feedback we've gotten from early users was not about the AI, the templates, or the workflow language. It was this: **connecting tools took too long**. General Input is amazing at building agents and workflows once it has the credentials it needs. Getting the credentials in was the bottleneck.
Register an OAuth app inside Salesforce. Paste an API token for HubSpot. Submit Microsoft Graph scopes to InfoSec for review. Reconnect Google Workspace because a token expired. A two-minute workflow sitting behind two hours of configuration.
## One-Click OAuth Integrations Now Live for Dozens of Providers
In the last week, we added **one-click integrations for dozens of providers**. The full Microsoft Suite, including Outlook, Teams, OneDrive, SharePoint, Excel, Planner, and To Do. The full Google Suite, including Gmail, Drive, Calendar, Sheets, Docs, Slides, Tasks, Meet, Analytics, and Forms. HubSpot, Salesforce, Reddit, Typeform, Attio, Airtable, Box, Dropbox, Basecamp, Calendly, GitHub, Linear, QuickBooks, Shopify, Stripe, and more.
The flow is now: pick a provider, click connect, authorize in the provider's own consent screen, done. No app registration inside your tenant. No tokens pasted into a form. No InfoSec review to get a basic integration into the hands of a builder.
The entire setup is three clicks: pick a provider, authorize, done. No app
registration, no tokens, no InfoSec review for basic integrations.
## BYO OAuth for Regulated Industries and Private Deployments
We chose to remove this friction for cloud customers, but stayed true to our values of enabling private deployment of AI by letting users bring their own OAuth in those cases. For regulated industries, enterprises with data residency requirements, or anyone still thinking about [last year's Salesloft Drift incident where a single compromised OAuth app exposed data in 700+ Salesforce instances](https://cloud.google.com/blog/topics/threat-intelligence/data-theft-salesforce-instances-via-salesloft-drift), bring-your-own OAuth is available for every provider where it matters. Register your own app, point General Input at it, and every token your users generate stays with you.
The security model underneath didn't regress. **Credentials never reach the AI.** They're encrypted at rest, injected into API calls at runtime, and stripped from request and response payloads before the model ever sees them. Every external call is audit-logged with the agent identity, the user on whose behalf it ran, and the exact endpoint and payload. [Okta is launching "Okta for AI Agents" on April 30](https://www.okta.com/newsroom/press-releases/showcase-2026/) with a blueprint built around the question "what can your agents connect to." This is our answer.
BYO OAuth is available for every provider. Register your own app, point
General Input at it, and every token stays in your tenant.
## Five Workflows You Can Build in an Afternoon
The point of shipping this was to stop making the connection step the thing that kills the workflow. Here are five workflows a RevOps team can stand up in an afternoon now that the integrations take minutes:
- **HubSpot Lead to Salesforce Opportunity Sync with Enrichment.** When a HubSpot contact hits MQL, enrich the company from the email domain, create the matching Salesforce opportunity, and post the handoff to a Microsoft Teams channel. Two CRMs stay in sync without a brittle Zap.
- **Stalled Salesforce Opportunity Alert to Microsoft Teams.** Daily scan for opportunities with no activity in 14 days. Enrich the account with recent news, generate an outreach angle with AI, and DM the Account Owner directly in Teams.
- **Weekly Pipeline Health Briefing to Outlook.** Every Monday at 7am, combine Salesforce pipeline and HubSpot funnel data into an AI-written briefing with pipeline coverage, stage movement, and at-risk deals. Emailed to leadership via Outlook.
- **Reddit Brand Mention to HubSpot Contact.** Monitor Reddit for mentions of your company or product. When a relevant post is found, enrich the author where possible and create a HubSpot contact with the Reddit thread logged as a note.
- **New Salesforce Lead Research Brief to OneDrive.** On every new Salesforce lead, run person and company enrichment, generate a structured research brief, save it to the account's OneDrive folder, and ping the Account Owner in Teams with the link.
Each one runs on connected tools, costs pennies per execution, and now takes minutes of setup instead of an afternoon.
## Why the Biggest Unlock Was Removing a Step
The useful lesson here isn't product design. It's that **the best feedback is often about what you should delete, not add**. Our early users didn't want a better integration configuration page. They wanted it to disappear. The floor that AI automation needs to cross isn't model quality. It's the friction between the idea and the first running workflow. Connecting tools was the friction that kept showing up. So we removed it.
If you tried General Input early on and got stuck at setup, come back. The
thing that stopped you is gone.
## FAQs
### Do I still need to register an OAuth app inside my tenant?
No, not for cloud customers. You click connect, authorize in the provider's own consent screen, and you're done. BYO OAuth is still available for regulated industries that need it.
### Which providers are one-click today?
The full Microsoft and Google suites, HubSpot, Salesforce, Reddit, Typeform, Attio, Airtable, Box, Dropbox, Basecamp, Calendly, GitHub, Linear, QuickBooks, Shopify, Stripe, and more. New providers ship every week.
### Is this less secure than registering my own OAuth app?
No. Credentials are encrypted at rest, injected at runtime, stripped before the model sees them, and every call is audit-logged. BYO OAuth is offered for tenants that want token ownership to stay in-house.
### I tried General Input early on and got stuck on integrations. Has anything changed?
Yes. Setup that used to take an afternoon now takes minutes. If integrations were the blocker, come back and try again.
---
# The AI Infrastructure Gap: What Ramp's Glass Reveals About Why Adoption Stalls
> 99% AI adoption still stalls without connected tools, shared workflows, and scheduled automations. Here's how to close the infrastructure gap.
Published: 2026-04-10. Last updated: 2026-05-07.
Ramp recently published [a detailed look at Glass](https://x.com/sebgoddijn/article/2042285915435937816), an internal AI productivity suite they built for every employee in the company. The premise is one that anyone who's tried to roll out AI across a team will recognize immediately: adoption isn't the hard part. **Making it stick is.**
They hit 99% AI tool adoption. Then they noticed most people were driving with the handbrake on. Terminal configurations, MCP setups, npm installs. The friction wasn't intellectual, it was environmental. And the few people who broke through had no way to share what they'd learned with everyone else.
## Why AI adoption stalls despite high usage
The AI discourse in 2026 is dominated by model capabilities. Which model reasons better, which one handles longer context, which one generates better code. But in practice, **the bottleneck at most organizations isn't the model. It's everything around it.**
[79% of enterprises face AI adoption challenges](https://writer.com/blog/enterprise-ai-adoption-2026/) despite massive investment. [Nearly two-thirds remain stuck in the pilot stage](https://hbr.org/2026/02/why-ai-adoption-stalls-according-to-industry-data). The pattern is consistent: companies buy licenses, employees sign up, and then nothing compounds. Each person uses AI in isolation. There's no shared infrastructure, no reusable workflows, no way for one person's breakthrough to become everyone else's baseline.
Ramp identified three things that actually move the needle: tools connected on day one so the AI has context, reusable skills that can be shared across the organization, and the product itself serving as the enablement layer rather than training programs or workshops.
## What Ramp's Glass actually solves
Glass is a full workspace (not a chat window) with split panes, persistent memory, scheduled automations, and a skill marketplace called Dojo where employees package workflows as shareable markdown files. Over 350 skills have been shared company-wide. An AI guide called the Sensei recommends the right skills based on your role and connected tools.
The architectural details matter. Memory is built automatically from your connected integrations. Automations run on cron schedules and post results to Slack. Everything connects via Okta SSO on day one. No configuration required from the employee.
Ramp's most important finding: the people who got the most value weren't the
ones who attended training sessions. They were the ones who installed a skill
on day one and immediately got a result.
## Why most companies can't build their own Glass
Glass is impressive. It's also the kind of project that required a dedicated team of six, months of development, and deep integration with internal systems like Okta SSO and proprietary APIs. Ramp is a $32B company with the engineering talent and budget to justify a custom build.
Most companies don't have Okta SSO, a dedicated IT team, or the ability to deploy custom internal software. A founder running a 10-person team, a RevOps manager at a Series A startup, a solo operator trying to automate their morning reporting. They all have the same problem Ramp's employees had. The model is good enough. The harness isn't set up. And they have no way to build the same solution.
## How to close the AI infrastructure gap without engineering
The infrastructure that makes Glass valuable isn't conceptually complex. It's three things: **connected integrations** so the AI has real data to work with, **reusable workflow templates** that encode what "good" looks like, and **scheduled automations** that run without human involvement.
General Input provides that scaffolding as a turnkey product. Connected tools on day one. Secure by default. Deployable on your infrastructure so data never leaves your walls. You connect your tools once, Google Workspace, Stripe, Slack, Notion, Salesforce, and [850+ others](https://generalinput.com). You describe a workflow in plain English and the AI builds it. Templates are installable in one click. Automations run on cron and deliver results to Slack, email, or wherever you need them. No IT team required.
A solo founder can set up a morning briefing that pulls yesterday's Stripe revenue, today's calendar, and flagged emails into one AI-compiled summary. A customer success manager can build a failed payment recovery pipeline that drafts personalized emails every morning. An ops lead can share a workflow template with their entire team the same way Ramp shares skills through Dojo.
Setup takes five minutes. Each run costs pennies. No engineering team
required.
## The floor rises for everyone
The most compelling line in Ramp's article is the last one: "We don't believe in lowering the ceiling. We believe in raising the floor."
That's the right frame. The goal isn't to simplify AI until it's useless. It's to make the full capability accessible without requiring every person to become a power user on their own. When the infrastructure does the heavy lifting, connecting tools, sharing workflows, maintaining context, people learn by doing. **The product is the enablement.**
Ramp proved the thesis internally. The challenge now is making that same infrastructure available to the other 99% of companies that can't build it themselves.
## FAQs
### Do I need an engineering team to set this up?
No. General Input is a turnkey product with connected integrations, reusable templates, and scheduled automations that work without IT involvement.
### Can I share workflows with my team like Ramp's Dojo skill marketplace?
Yes. Workflows can be shared across an organization, and templates package them as installable building blocks anyone on your team can adopt in one click.
### How long does setup actually take?
About five minutes. Connect the tools, pick a template or describe the workflow, then schedule it. Each run costs pennies.
### How does General Input handle data privacy compared to general-purpose AI tools?
It's deployable on your own infrastructure for teams that need data to stay inside their walls. Credentials stay encrypted at rest and never reach the model.
---
# BYO Model Deployment: AI Agents That Never Phone Home
> Run AI agents on your own infrastructure. BYO model deployment supports Bedrock, Vertex, Azure, Ollama, and any OpenAI-compatible endpoint.
Published: 2026-04-06. Last updated: 2026-05-07.
Google dropped [Gemma 4](https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/) last week. A family of open models purpose-built for agentic workflows: multi-step planning, autonomous action, tool use. The 31B dense model [ranks #3 on Arena AI](https://developers.googleblog.com/bring-state-of-the-art-agentic-skills-to-the-edge-with-gemma-4/) at 1452 Elo, outperforming models twenty times its size on reasoning benchmarks. Apache 2.0 license.
Meanwhile, a developer got [Qwen3.5-397B running on a 48GB MacBook Pro](https://www.cultofmac.com/news/data-center-ai-model-on-macbook). A model that normally requires a server rack. Any Apple Silicon Mac with 16GB of RAM can [run useful local models](https://apxml.com/posts/best-local-llm-apple-silicon-mac) today via Ollama or LM Studio.
Let me be honest: none of these are replacing Claude or GPT-5 for complex agent workflows right now. If you need an AI to orchestrate five APIs, reason about edge cases, and handle errors gracefully, you still want a frontier model. But the trajectory is undeniable. **The gap between local and frontier shrinks every quarter.** We're trending toward a world where everyone deploys agents from their laptop.
The infrastructure should be ready before the models are.
## Why your AI provider choice matters
Most workflow automation tools process your data on their servers using their AI provider. Your CRM records, payment data, customer information, and internal documents flow through infrastructure you don't control.
That made sense when running AI required a data center. It makes less sense every month. Companies already run their own Claude on AWS Bedrock, Gemini on Google Vertex, and OpenAI on Azure, governed by their IAM policies, logged by their CloudTrail, contained within their VPC. [73% of enterprises](https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html) cite data privacy as their top AI risk. The answer isn't to avoid AI. It's to run it on your own terms.
## What BYO model deployment means for General Input
I added BYO model deployment to General Input this weekend. Instead of using the built-in AI, you point the agent at your own model endpoint. The workflow engine handles scheduling, integration orchestration, credential management, and the audit trail. The AI inference happens wherever you point it.
Today the practical use is **corporate cloud deployments**. Claude on Bedrock. Gemini on Vertex. OpenAI on Azure. Any OpenAI-compatible endpoint. If your security team already spent months getting a model approved and deployed inside your VPC, you shouldn't have to re-litigate that conversation every time you want to automate a workflow.
The infrastructure is model-agnostic. It doesn't care whether the endpoint is in us-east-1 or running on your desk. When local models are good enough for production agent work, and based on the last two years of progress, that's a when, not an if, the same BYO setup works with Ollama on your Mac. And the inference cost drops to zero. No per-token charges from a cloud provider. Just your hardware and electricity.
BYO deployment works with any OpenAI-compatible endpoint. That covers Bedrock,
Vertex, Azure OpenAI, Ollama, LM Studio, vLLM, and anything else that speaks
the same protocol.
## How credentials and data stay private
BYO model deployment sits on top of an architecture designed for this from day one.
**Credentials never reach the AI.** API keys and OAuth tokens are encrypted at rest, injected at runtime when the agent calls an external service, and stripped from the context before data reaches the model. The agent orchestrates your tools but never holds the keys.
**Every data access is logged.** When the agent reads from your CRM, pulls transactions from Stripe, or sends a Slack message, each access is recorded in an audit trail.
**The model is yours.** With BYO deployment, the AI inference itself runs on infrastructure you control. Your data enters your model and comes back as instructions. It never touches a third-party AI provider.
Your data never leaves your walls, simplifying HIPAA, SOC 2, GDPR, and other
privacy-sensitive deployments.
## Where local AI agents are heading
Gemma 4's 2B and 4B models are explicitly designed for on-device deployment. They handle multi-step planning and autonomous action at sizes that fit on a phone. The [31B model runs at chat speed](https://siliconangle.com/2026/04/02/googles-new-gemma-4-models-bring-complex-reasoning-skills-low-power-devices/) on a Mac with 48GB of unified memory. A year ago, running an AI agent locally was a science project. Today it's getting close to practical.
BYO model deployment is the bridge. It works with enterprise cloud deployments today, and it'll work with whatever comes next. When it does, you get **private data, zero inference cost, and no vendor lock-in**, all on hardware you already own.
We're not there yet. But every quarter, we get closer.
## FAQs
### What does BYO model deployment let me do?
Point your General Input agents at any OpenAI-compatible endpoint. Claude on AWS Bedrock, Gemini on Vertex, OpenAI on Azure, or a local Ollama install on your own hardware.
### Will my data ever leave my infrastructure?
With BYO model deployment, no. Inference runs on the endpoint you choose, and credentials never reach the model regardless of where it's hosted.
### Is this useful today, or only for some future where local models are frontier?
Useful today. The most common use is corporate cloud deployments: running a model your security team has already approved inside your VPC. The same setup works tomorrow when local models are good enough.
### How does this help with HIPAA, SOC 2, or GDPR?
Your data stays inside your infrastructure boundary, so you don't have to re-litigate every workflow with security and compliance. The model is yours, the audit trail is yours, the keys never cross a perimeter you don't own.
---
# How to Build a Self-Maintaining CRM with Cal.com, Attio, and Granola
> Wire Cal.com, Attio, and Granola into three workflows that enrich leads on booking, brief reps before calls, and log meeting notes automatically.
Published: 2026-03-30. Last updated: 2026-05-07.
You have a call in 10 minutes. You open Attio, search for the prospect, find a half-empty record from three months ago. No company info, no LinkedIn, no notes from the last conversation. You Google them, skim their company's About page, and hope you remember what you discussed last time.
After the call, you have five minutes before the next one. You should update Attio with what was discussed, the next steps you agreed to, and the fact that they're evaluating a competitor. Instead, you jump straight into the next call. The notes stay in your meeting app. The CRM stays stale.
This is [how 79% of opportunity data goes missing](https://devrev.ai/blog/sales-reps-hate-crm). Not bad data. **Missing data.**
## Three gaps in every sales rep's meeting cycle
CRM decay isn't one problem. It's three.
**At booking time**, new leads arrive via Cal.com but nobody creates the CRM record until the meeting is imminent, if at all. Leads slip through. Duplicates pile up.
**Before the meeting**, reps waste time assembling context. [Sales professionals spend roughly a third of their time](https://www.docket.io/blog/preserve-sales-institutional-knowledge) locating information before calls. They're checking LinkedIn, Googling the company, searching old email threads. All because the CRM record doesn't have what they need.
**After the meeting**, reps are supposed to log notes. [68% of sales professionals](https://prospeo.io/s/sales-productivity-statistics) say this is their most time-consuming task. After back-to-back calls, it gets skipped.
The tools exist to fix all three. Cal.com knows who's booking. Attio holds the records. Granola captures what was discussed. Three workflows connect them into a loop.
## Why Granola changes the equation
Granola captures meeting notes by hooking into your system audio. No bot joins your call, no awkward "this meeting is being recorded by an AI" moment. You jot down a few keywords during the call, and afterward Granola's AI fleshes them out into structured notes with decisions, action items, and key quotes.
It just [launched a Personal API](https://techcrunch.com/2026/03/25/granola-raises-125m-hits-1-5b-valuation-as-it-expands-from-meeting-notetaker-to-enterprise-ai-app/) that makes meeting notes programmatically accessible. Meeting content was the last piece locked in a silo. Cal.com already has webhooks, Attio already has an API. Now Granola is a data source too. So I integrated it on General Input and built the self-maintaining CRM I actually use to run my own sales process. Here's the setup.
## How to set up three automated CRM workflows
### Workflow 1: Cal.com booking to Attio lead with enrichment
When a demo meeting is booked via Cal.com, the workflow takes the guest's email and checks Attio for an existing contact. If they're already in the CRM, it skips creation. If they're new, it creates the lead. Then it researches the guest. Company Enrichment pulls industry, size, funding, and tech stack. Person Enrichment adds job title, career history, and social profiles. Research notes get added to their Attio record. The lead is ready before the rep thinks about it.
### Workflow 2: Pre-meeting sales brief via Slack
One hour before the meeting, the rep gets a Slack DM with a briefing. It pulls the Attio contact record (including the enrichment notes from Workflow 1 and any past meeting notes from Workflow 3), searches for recent news about the company, and packages it all into a concise brief. The rep walks into the call informed without opening a single tab.
### Workflow 3: Granola meeting notes synced to Attio
After the meeting, the workflow pulls the enhanced notes from Granola, matches the attendee to their Attio record, and adds the meeting content. What was discussed, decisions made, next steps. The CRM record stays current. When the same contact books another meeting weeks later, Workflow 2 will include these notes in the next pre-meeting brief.
The three workflows form a cycle. Each one feeds the others. The CRM enriches
on booking, briefs the rep before the call, and updates itself after. The rep
never opens Attio.
Setup: about 5 minutes per workflow. Cost: pennies per run.
## The compounding effect over time
The real value shows up on the third, fourth, fifth meeting with the same contact. Each conversation's notes accumulate in Attio. Each pre-meeting brief gets richer because it draws on more history. The CRM becomes a **genuine knowledge base**, not a graveyard of "call completed" entries.
When a rep hands off an account or leaves the company ([sales turnover runs ~35% annually](https://www.xactlycorp.com/blog/motivation/sales-turnover-statistics)), the context stays in the system. The next rep gets the same briefings, the same history, the same enrichment. Nothing walks out the door.
This pattern works with any booking tool, CRM, and meeting-notes service that
exposes an API. Cal.com, Attio, and Granola are the example stack, not a
requirement.
## FAQs
### Do I need all three of Cal.com, Attio, and Granola to make this work?
They're the example stack, but the same pattern works with any booking tool, CRM, and meeting-notes service that exposes data via API or webhooks.
### Does Granola require a bot to join my calls?
No. Granola hooks into your system audio and turns rough notes into structured meeting summaries via its Personal API. There's no awkward bot in the meeting.
### Will the workflow create duplicates if a contact is already in Attio?
No. Workflow 1 checks Attio for an existing contact by email before creating a new lead, then layers enrichment onto whichever record matches.
### How long does it take to set up the three-workflow loop?
About five minutes per workflow, or roughly fifteen minutes for the full loop. Each run costs pennies.
---
# I Understand Why Leaders Are Nervous About Hooking Their Business Data Into AI
> Most AI tools store API keys in plain text. The LiteLLM hack proved why that's dangerous. Learn how credential isolation keeps your keys away from the model.
Published: 2026-03-25. Last updated: 2026-05-07.
When you connect a tool like OpenClaw to your internal systems, you hand it your passwords and API keys. Most tools store those in plain text. Like writing your bank password on a sticky note and leaving it on your desk. Anyone who gets access to the machine can read them.
That's exactly what happened this week.
## The LiteLLM supply chain attack, explained
[LiteLLM](https://www.theregister.com/2026/03/24/trivy_compromise_litellm/) is a popular open-source tool that connects AI agents to models like Claude and ChatGPT. If you're using OpenClaw or similar tools, there's a good chance LiteLLM is running in the background routing your requests. It has [95 million downloads per month](https://www.bleepingcomputer.com/news/security/popular-litellm-pypi-package-compromised-in-teampcp-supply-chain-attack/).
Attackers first compromised a security scanner called Trivy that LiteLLM used in its build process. That gave them the keys to push a malicious software update directly to PyPI, where developers download Python packages. Anyone who installed the update had **every password and API key on their machine silently copied and sent to the attackers**. No warning. Just installing the update was enough.
The malicious versions were live for about three hours before they were caught
and removed. That's all it takes.
## Why AI automation makes API key security critical
If your passwords are stored in plain text where the AI can see them, **any breach in the chain exposes everything**. It doesn't matter how good the AI model is or how useful the automation is. The weakest link is wherever the credentials live.
Most AI automation tools work like this: you connect your systems, the tool stores your API keys, and when a workflow runs it passes those keys to the AI model so it can make requests on your behalf. Your credentials end up sitting in the model's memory, flowing through infrastructure you don't control.
LiteLLM showed what happens when one link in that chain gets compromised.
## How credential isolation keeps your keys away from the AI
The approach I took with General Input comes down to three things:
### Limit access with scopes
You control exactly what the AI can and can't do in each system. Read-only access to your CRM? Fine. No access to billing? Done. The AI only gets the permissions you explicitly grant.
### Credentials encrypted at rest, never shown to the AI
The AI decides what to do. A separate execution engine handles the actual login to your tools at the moment it's needed, then strips the credentials before the AI ever sees the response. Think of it like giving someone driving directions without handing them your car keys.
This is the key architectural difference. The model never touches your
credentials. It describes what it wants to do, and a sandboxed execution layer
handles authentication separately.
### Full audit log with a reason for every access
Every API call, every field, every record. Logged with an explanation of why it was accessed. Not a summary, not "the workflow ran successfully." An actual log of what was touched, when, by which step, and for what purpose.
For teams with strict compliance requirements: General Input can be
self-hosted. Your cloud, your AI model, your credentials, your audit logs.
Nothing crosses a boundary you don't own.
## The fear is justified. The architecture should account for it.
Leaders are right to be cautious about connecting their business systems to AI. The question isn't whether to be nervous. It's whether the system you're using was built with that concern in mind from the start.
**Sources:**
- [LiteLLM supply chain attack (The Register)](https://www.theregister.com/2026/03/24/trivy_compromise_litellm/)
- [LiteLLM supply chain attack (BleepingComputer)](https://www.bleepingcomputer.com/news/security/popular-litellm-pypi-package-compromised-in-teampcp-supply-chain-attack/)
## FAQs
### Can the AI model see my API keys or OAuth tokens?
No. Credentials are encrypted at rest, injected into the sandboxed step that calls your tool, and stripped from request and response payloads before the model ever sees them.
### Was General Input affected by the LiteLLM supply chain attack?
No. Our credential-handling architecture doesn't pass tokens through LiteLLM or any model proxy that holds plaintext keys.
### What does the audit trail actually contain?
Every API call, the field or record touched, the step and workflow that ran it, the user identity behind the run, and the reason. Not just "workflow succeeded."
### Can I self-host General Input for compliance reasons?
Yes. For teams with strict data residency or compliance requirements, the platform deploys in your own cloud so credentials, audit logs, and inference all stay inside your boundary.
---
# Build Your Own Revenue Agent: What Zuckerberg's AI Assistant Means for Ops Teams
> Zuckerberg built an AI agent to skip the layers between him and data. Here's how ops teams can build a revenue agent that delivers Monday's numbers.
Published: 2026-03-23. Last updated: 2026-05-07.
Mark Zuckerberg just built himself an AI agent. Not for Meta's 3 billion users. For himself. According to the [Wall Street Journal](https://www.wsj.com/tech/ai/mark-zuckerberg-is-building-an-ai-agent-to-help-him-be-ceo-eddab2d5?st=TUkpQJ&reflink=desktopwebshare_permalink), the agent helps him "retrieve answers he would typically have to go through layers of people to get."
That sentence should resonate with anyone who runs a team.
## What the Monday morning data pull actually looks like
Most managers don't have layers of people between them and the data. **They are the layer.** They're the person everyone pings when they need a number, and they're the person who has to go get it from four different systems.
A typical Monday morning looks something like this: log into Salesforce, pull the pipeline by stage, export to a spreadsheet. Open Stripe, check collected revenue and failed payments. Open Google Sheets, paste everything into the weekly report template you built six months ago. The one held together by formulas you're afraid to touch.
One ops manager described inheriting a weekly revenue report with **eight source inputs** that took four hours to run on a good day. Leadership spent about five minutes reading the actual insights.
Salesforce's State of Sales report says reps spend 70% of their time on
non-selling activities. But the person compiling the numbers for those reps?
They're spending Monday morning being the human middleware between Salesforce
and a Slack message.
## Why Zuckerberg's AI agent matters beyond Meta
The interesting thing about Zuckerberg's CEO agent isn't the technology. It's the framing. He didn't hire another chief of staff. He didn't build a dashboard. He built something that goes and gets the answer for him, from whatever system it lives in, and brings it back.
Meta's employees are doing the same thing internally. A tool called Second Brain, built by a Meta employee on top of Claude, is described as an "AI chief of staff" that indexes documents and answers questions across projects. Employees have personal agents that talk to their colleagues' agents on their behalf.
This is happening at a company with 78,000 employees and a $1.5 trillion market cap. **62% of organizations are now experimenting with AI agents**, and the market is projected to hit $10.9 billion this year. The question isn't whether AI agents are ready for business operations. It's whether you're going to keep being the human version of one.
## How to build a revenue agent in 3 minutes
You don't need Meta's engineering team to build something useful. Here's a concrete example.
I gave a specialist the prompt: "Every weekday at 7am, answer these questions and email me the answers: How much did we collect yesterday? (Stripe) Are any payments failing right now and who owns those accounts? (Stripe + Salesforce) What deals are closing this week and are any overdue? (Salesforce) Who am I meeting with today and what's their deal status? (Google Calendar + Salesforce) Did anyone cancel or downgrade yesterday? (Stripe) Keep it short. Bold the numbers. Flag anything urgent at the top."
**That took about 3 minutes to set up. It costs a few cents per run.** Now every morning, the answers land in my inbox before I open anything else.
The difference between this and a dashboard: nobody has to go look at it. It
comes to you, written in plain English, not charts that require
interpretation. And it pulls from three different systems that don't natively
talk to each other.
## Five revenue workflows worth automating
Once you see how this works for the Monday brief, the natural question is: what else am I compiling by hand?
### Deal risk detector
Every day, scan Salesforce for deals past their expected close date or with no activity logged in the last 14 days. Cross-reference Stripe for any failed payments on those accounts. Send a risk digest to the sales manager in Slack with deal names, owners, and the specific risk signal.
### Forecast vs. actuals reconciliation
At the end of each month, pull the forecasted revenue from Salesforce and compare it to what Stripe actually collected. Calculate variance by rep and by segment. Deliver the reconciliation as a Google Sheet and post a summary to Slack.
### CRM data hygiene scorecard
Every Friday, audit Salesforce deals for missing close dates, blank amount fields, and empty next-step notes. Calculate a data completeness percentage per rep. Send each rep their score via Slack with the specific fields to fix.
### Account 360 briefing
Before any customer meeting, pull the Salesforce deal record, Stripe payment history, and recent Zendesk tickets. Compile a one-page briefing in Google Docs and share the link in Slack.
### New deal auto-enrichment
When a new opportunity is created in Salesforce, enrich the company using Company Enrichment and search for recent news with Internet Search. Add firmographic data and a news summary to the deal description so the rep has context before their first call.
## Stop being the human middleware
Zuckerberg's insight wasn't complicated. He wanted information faster and he realized the bottleneck wasn't the data. It was the path between him and the data. He built something to shorten that path.
If you run a team, you probably recognize that bottleneck. **You are the path.** Every Monday morning, every end-of-month close, every "hey can you pull the numbers on..." ping in Slack. The information exists. It just lives in five different tools and someone has to go get it.
That someone doesn't have to be you anymore.
## FAQs
### How long does it take to set up the morning revenue brief?
About three minutes once your tools are connected. Cost is pennies per run.
### Which systems can the revenue agent pull from?
Salesforce, Stripe, HubSpot, Google Calendar, Zendesk, Postgres, and dozens more. Anything General Input integrates with.
### How is this different from a dashboard?
A dashboard waits for you to open it. A revenue agent comes to you in plain English: bolded numbers, urgent items flagged, every weekday at 7am, in your inbox or Slack.
### Can I make different agents for different roles on my team?
Yes. Each workflow is its own agent with its own data sources and delivery channel. One for the CFO's collections digest, another for the sales manager's deal-risk alerts.
---
# How to Automate Meeting Prep When Your Customer Data Lives in 5 Different Systems
> Automate pre-meeting briefs by pulling CRM, billing, email, and product data into one doc. Save 15 minutes per call with zero manual tab-switching.
Published: 2026-03-20. Last updated: 2026-05-07.
You have a customer call in 30 minutes. So you open Attio to check the deal stage and last notes. Open Stripe to see if they're current on payments. Search Gmail for the last email thread. Open Google Docs for notes from the previous meeting. Then run a Postgres query to check their product usage, because you want to know if they've actually been active before you get on the call.
**Five tabs. Ten to fifteen minutes. And you do this three to four times a day.**
## Why pre-meeting prep eats your schedule
If you're a B2B SaaS founder with 10 to 50 customers and you still handle most calls yourself, this ritual is deeply familiar. No single step is hard. Pulling a CRM record takes 30 seconds. Checking Stripe takes a minute. Running a usage query takes another couple minutes. But stringing them all together before every call, every day, is the kind of work that quietly eats your schedule.
Research shows founders spend only **32% of their time on growth activities**, with 68% consumed by operational tasks. The tab-switching meeting prep ritual is a textbook example. You're not doing anything strategic. You're assembling context that should already be assembled.
## The real cost is context switching, not the individual steps
Asana's research found that digital workers toggle between apps **1,200 times a day**. Context switching can cost up to 40% of productive time. Meeting prep is the most common version of this problem. Not because any individual step is hard, but because you're stitching together data from four or five systems by hand, every single time.
Ramp's March 2026 SaaS spending report shows nearly 1 in 4 businesses are
paying for AI tools. But the biggest unlock isn't the AI itself. It's
eliminating the small, repetitive cross-system tasks that eat your day one tab
at a time.
## How to automate meeting prep with AI
On General Input, the setup is one prompt to a specialist: "Before every meeting on my Google Calendar with an external attendee, pull their Attio deal record, Stripe payment history and current MRR, the last 5 Gmail threads, and query our PostgreSQL database for their usage stats over the past 30 days. Compile a one-page brief in Google Docs and send me the link in Slack 30 minutes before the call."
**Three minutes to set up. A few pennies per run.** Now you walk into every call knowing the deal stage, payment status, last conversation, and whether they've actually been using the product. No tabs, no scrambling. The brief is just there, waiting in Slack, 30 minutes before the call starts.
Start with your highest-stakes meetings. Once the workflow is running, extend
it to all external calls on your calendar.
## Five automated meeting workflows for founders
The pre-meeting brief is the starting point. Here are four more workflows that extend the same idea:
### 1. Pre-meeting customer brief
Before each customer meeting, pull their Attio deal record, Stripe billing history, recent Gmail threads, and product usage stats from PostgreSQL. Compile a one-page brief in Google Docs and send the link via Slack.
### 2. Monday meeting prep digest
Every Monday morning, scan the week's Google Calendar for external meetings. Cross-reference each company against Attio deal stage and Stripe payment status. Deliver a prep digest via email.
### 3. Post-meeting CRM and follow-up sync
After a meeting ends, use AI to extract action items from notes in Google Docs, create Google Tasks, update the Attio deal record, and draft a follow-up email in Gmail.
### 4. Customer payment context alert
When a customer meeting appears on Google Calendar, check Stripe for failed payments, overdue invoices, or upcoming renewals. Send a heads-up via Slack before the meeting.
### 5. New lead meeting research brief
When a new external meeting is added to Google Calendar, research the attendee's company with internet search and company enrichment. Deliver a background brief via Slack.
## Walk into every call prepared, without doing the prep
The meeting itself is the valuable part. The prep is just overhead. When the overhead runs itself, you get back **an hour or more per day** that you were spending on tab-switching and data assembly. That's time you can spend on the conversation, the follow-up, or the product work that actually moves the business forward.
## FAQs
### Which systems can the pre-meeting brief pull from?
Calendar, CRM (Attio, Salesforce, HubSpot), billing (Stripe), email (Gmail or Outlook), product databases (Postgres or anything via API), and docs. Anything you currently open as a tab.
### How early before the meeting does the brief arrive?
Configurable. Most founders set it to 30 minutes before each external meeting, but you can change the timing in the workflow.
### Where does the brief get delivered?
Wherever you read first thing: Slack DM, email, or a Google Doc link. You pick.
### Does this work for back-to-back days of customer calls?
Yes. The workflow runs against every external meeting on your calendar automatically, so a 10-meeting day still costs you no manual prep time.
---
# You Don't Need to Be Technical to Win With AI Agents Anymore
> AI agents now work for non-technical operators. Describe a workflow in plain English and have it running in minutes, no code required.
Published: 2026-03-19. Last updated: 2026-05-07.
Jason Lemkin laid it out on 20VC this month. The AI era has had three phases.
**Phase 1 was too technical.** Only engineers could touch it. If you wanted to automate something, you needed to write code, manage APIs, and deal with infrastructure. The people who understood the broken processes weren't the ones who could fix them.
**Phase 2 was prompt engineering.** The tools opened up a bit, but you still had to torture them into doing what you wanted. You could make things work, but it required a specific kind of patience and fluency that most operators didn't have time to develop.
**Phase 3 is now.** In Lemkin's words: "Mere ordinarily smart generalists can make these tools do magical things."
## Why this shift matters for operations teams
The most important thing about this phase isn't the AI getting better. It's **who gets to use it**.
The person who understands a broken process best is the person stuck inside it. The CSM who manually assembles QBR decks every quarter. The sales rep who Googles prospects one by one. The founder who opens five tabs before every customer call. These people have always known exactly what should be automated. They just couldn't do anything about it.
That's changing. The gap between "I know what needs to happen" and "I can make it happen" is collapsing. You don't need to know how to write code to describe a workflow in plain language and have it run.
Pick the most repetitive part of your week. Describe it in plain English. Have
it running in minutes. That's the entire workflow.
## "Pick yourself off the ground. This is your time."
Lemkin's point isn't just observational. It's a call to action. The people who have been waiting for automation tools to get easy enough are out of excuses. The tools are there. The barrier isn't technical skill anymore. It's just deciding to do it.
The operators, the generalists, the people who have been manually holding processes together with spreadsheets and willpower are the ones best positioned to build automations that actually work. They know the edge cases. They know which steps get skipped. They know what breaks on Fridays at 4pm.
## What changes when the operator builds the automation
When technical skill was the bottleneck, automation projects went through a translation layer. An operator would describe what they needed. A developer would interpret it. Something would get built. It would be close but not quite right. Iteration cycles were long. Half the projects died in the gap between "what I asked for" and "what got built."
Remove the translation layer and the feedback loop compresses to minutes. The
person who knows the process builds the automation, tests it against reality,
and adjusts. No spec documents. No sprint planning. No waiting.
**Nobody understands a broken process better than the person stuck inside it.** Now they can actually do something about it.
## FAQs
### Do I need any coding background to build an AI workflow?
No. The current generation of tools accepts plain-language descriptions and produces a working automation. The bottleneck is no longer technical skill. It's deciding which workflow to build first.
### What kinds of operators are getting the most leverage right now?
People closest to the broken process: CSMs assembling QBRs, SDRs prospecting, founders prepping for customer calls. They know the edge cases and can iterate in minutes instead of waiting for an engineering sprint.
### Where should a non-technical operator start?
Pick the most repetitive part of your week, describe it in plain English, and let the AI build the first version. Adjust against reality from there.
### How is this different from no-code tools I tried before?
No-code still required configuring each step through dropdowns and nodes. Now you describe the outcome and the AI does the wiring. You read and review, not build.
---
# How to Use Google Ratings as a Sales Prospecting Signal for Local Businesses
> Low Google ratings reveal operational problems your software can solve. Turn review data into a qualified prospect list automatically.
Published: 2026-03-18. Last updated: 2026-05-07.
Sales reps selling to local businesses all do the same thing. Google "HVAC companies Nashville." Click through results one by one. Check the rating. Copy the name, phone number, and address into a spreadsheet. Move to the next metro. Repeat every week. SPOTIO's 2026 State of Field Sales Survey found that **33% of field sales teams still aren't using AI at all**, and reps continue spending 70% of their time on non-selling activities. Almost half of that time is just finding someone to call.
## What sales prospecting looks like when you sell to local businesses
If you're an SDR at a vertical SaaS company selling field service management, reputation management, or scheduling software, your target customers are HVAC contractors, plumbers, electricians, and similar local service businesses. Your territory might cover a dozen metros.
The typical workflow: open Google Maps, search for your target industry in a metro, scroll through results, and evaluate each business. You're looking at ratings, review counts, and sometimes reading individual reviews to gauge whether the business has problems your software could fix. Then you copy the promising ones into a spreadsheet with their contact info.
This takes hours. And you do it every week because new businesses appear, ratings change, and you need fresh prospects. **42% of reps say prospecting is the hardest part of their job.** When you look at how the sausage gets made, it's not hard to see why.
## Why Google ratings are an underused prospecting signal
Here's what most reps miss: Google ratings are one of the strongest indicators of whether a local business needs help.
A 2.9-star HVAC company with 85 reviews isn't just a bad business. Read the reviews. They're full of "missed my appointment," "never called back," "had to chase them for an invoice." Those are **operational problems**. If you sell software that fixes scheduling, dispatch, or customer communication, that business is practically raising their hand.
Below 4.0 stars, Google starts filtering businesses out of local search
results. Only 3% of consumers will trust a business under 3 stars. These
companies are losing customers in real time, and most of them know it.
That changes your cold outreach from "hey, want to buy some software?" to "I noticed your customers are mentioning scheduling issues, and that's exactly what we fix."
## How to automate your prospect list with Google ratings
Instead of clicking through Google results by hand, you can set this up as a workflow that runs itself. On General Input, you tell a specialist: "Every Monday, take my list of target metro areas, search for HVAC companies in each one, filter for businesses rated below 4.0 stars, and append new results to my Google Sheet. Deduplicate by Place ID so I never see the same business twice. Include name, address, rating, review count, phone number, and the date they were added."
**Setup takes a few minutes. Costs a few pennies per search.** Places data and Google Sheets are built into the platform, so there are no API keys to configure and no third-party accounts to manage. Every Monday you wake up to a fresh list of qualified prospects who are already signaling that they need what you sell.
This works for any local service vertical. HVAC, plumbing, electrical,
roofing, dental, restaurants. Anywhere Google ratings reflect operational
performance.
## Five workflows for sales reps prospecting local businesses
Beyond the core prospect finder, there are several workflows that build on this approach:
1. **Local Business Prospect Finder** -- search for businesses in a specific industry and metro, filter by Google rating, and output a prospect list to Google Sheets with name, rating, review count, phone, and website.
2. **Weekly Multi-Metro Prospect Builder** -- run the prospect search across multiple metros on a schedule, append new results to a master sheet, and deduplicate against businesses already found.
3. **Low-Rating Outreach Brief Generator** -- for each low-rated prospect, pull recent review themes and generate a personalized outreach brief highlighting specific pain points their software solves.
4. **New Business Alert Monitor** -- watch for newly listed businesses in target categories and metros, alert the rep in Slack with business details.
5. **Competitor Customer Review Tracker** -- monitor reviews for mentions of competitor products, flag businesses expressing frustration, and add them to the prospect sheet.
## From cold prospecting to signal-based selling
The difference between cold-calling a list from a data vendor and reaching out to a business whose own customers are publicly describing the exact problem you solve is enormous. One is a numbers game. The other is **pattern recognition at scale**.
When the prospecting itself runs automatically, reps can spend their time on the part of the job that actually requires a human: reading the reviews, crafting a relevant message, and having the conversation.
## FAQs
### Do I need a Google API key or paid data vendor for this?
No. Places data and Google Sheets are built into the platform. No API keys to configure and no third-party accounts to manage.
### How does the workflow avoid showing me the same business twice?
It deduplicates by Place ID against your master sheet, so each Monday you only see businesses that are new since the last run.
### What rating threshold should I filter on?
Below 4.0 is the inflection point. Google starts deprioritizing those listings in local search, and reviews tend to surface operational complaints. You can adjust the threshold to fit your ICP.
### Can I extend this beyond HVAC and into other local-business verticals?
Yes. The same pattern works for plumbers, electricians, roofers, dentists, restaurants, or any local business where Google ratings reflect operational performance.
---
# Introducing General Input: Workflow Automation Written by AI
> General Input lets you describe workflows in plain English. An AI agent builds them, you review and run them. No code, no flowcharts.
Published: 2026-03-17. Last updated: 2026-05-07.
I started my career building workflow automations in tools like Microsoft Power Automate and Zapier, and later moved into building custom software. That experience always left me feeling like there was a tradeoff. Traditional automation tools were more accessible, but often tedious to configure, edit, and maintain. Custom software was far more flexible, but usually required **engineering resources and expertise** that is in short supply.
## The tradeoff between accessible and powerful automation
No-code automation tools solved a real problem. They let non-technical people build workflows without writing code. But anyone who has spent time in a drag-and-drop builder knows the frustration: connecting nodes, configuring each step, handling edge cases through menus and dropdowns, and debugging by clicking through a flowchart trying to figure out where things went wrong.
The best automations have always looked more like software than flowcharts. They have conditional logic, error handling, data transformations, and decision points that are **easier to express in language than in a visual builder**. But software requires engineers, and engineering time is expensive and slow. Most teams can't justify building custom software for everyday operational workflows.
The gap has always been the same. The people who understand broken processes
aren't the ones with the technical skills to fix them.
## How AI closes the automation gap
That's changing. AI makes it possible to describe what you need in plain language and get working software back. Not a dumbed-down version. Not a visual approximation. Actual working automation that reads like a document, that you can review and modify, and that runs on a schedule or in response to events.
General Input is built from the ground up for this. You describe what you need to a specialist, our AI agent: what to check, which tools to use, what should happen next, and what decisions matter. The specialist turns that into a working automation.
## What AI-written workflow automation looks like in practice
One example: a workflow that looks across open Salesforce opportunities, identifies deals that have been stuck in the same stage too long, researches recent company news, summarizes what may be affecting each account, generates a polished PDF report for each rep, and sends it directly in Slack.
**Built in under 5 minutes for a few pennies per run.** That's a workflow that would take days to build in a traditional automation tool and weeks if you were building it as custom software. The AI handles the implementation. You handle the intent.
Start with the workflow you dread most each week. Describe it to a specialist
in plain English and have it running in minutes, not days.
## Ready-to-use workflow templates across sales, ops, and customer success
Beyond building custom workflows from scratch, General Input has dozens of ready-to-use workflow templates. Connect your tools and activate them instantly. Templates cover sales prospecting, pipeline management, customer success monitoring, operations reporting, and more.
Each template is a real workflow, not a skeleton. They come with the logic, the integrations, and the decision points already wired up. You can use them as-is or tell a specialist to modify them for your specific needs.
## What's next for General Input
General Input is currently in beta. If you're an operator, RevOps leader, sales leader, or automation-minded team, we'd love to hear from you and learn what workflows matter most to your team.
We'll be sharing workflow demos, use cases, and behind-the-scenes builds regularly. The goal is simple: the people who understand broken processes should be the ones who fix them, and they shouldn't need an engineering team to do it.
## FAQs
### How is this different from Zapier or Power Automate?
Workflows are written in plain English by an AI agent rather than configured node by node in a flowchart builder. They read like documents you can review and modify, and they handle conditional logic, data transformation, and decision points natively.
### Do I need to write code to build a workflow?
No. You describe what you need to a specialist in plain English. The specialist produces a working automation you can review, modify, and run on a schedule or in response to events.
### Are there pre-built workflows I can use?
Yes. Dozens of templates cover sales prospecting, pipeline management, customer success, operations reporting, and more. Each is a real workflow with logic and integrations already wired up.
### Who is General Input built for?
Operators, RevOps leaders, sales leaders, and any team where the people closest to a broken process should be the ones fixing it, without needing engineers.
---
# General Input changelog
> Everything new in General Input, newest first, and what each change means for your work.
## Alchemy, Northflank, DigiCert, and 47 more integrations
> Fifty integrations joined the catalog, from Alchemy, Northflank, and DigiCert to Synthflow, Veryfi, and WATI, each with documented operations and a key check.
Oct 14, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-14-alchemy-northflank-digicert-and-more
Fifty more integrations are live, each connected with an API key: developer platforms and infrastructure, AI agents and voice, document and media processing, web data, market data, publishing, messaging, and team tools. Find them in the [integrations catalog](/apps), including [Alchemy](/apps/alchemy), [Northflank](/apps/northflank), and [DigiCert](/apps/digicert).
### What was added
- **Developer tools:** Alchemy, Northflank, Turso, Codemagic, Codacy, Sauce Labs, SwaggerHub, Mintlify, npm, InfluxDB Cloud, DigiCert, and Hostinger.
- **AI and agents:** Manus, Flowise, Synthflow, Supermemory, Skyvern, Meshy, and You.com.
- **Documents and media:** Nanonets, Veryfi, Photoroom, DocRaptor, Parseur, Flatfile, ScreenshotOne, and Creatomate.
- **Web and market data:** ZenRows, Scrapfly, Browse AI, ScrapeGraphAI, Benzinga, Nasdaq Data Link, and Tripadvisor.
- **Publishing and messaging:** Typefully, Ayrshare, Hashnode, DEV, Landbot, WATI, and Pushover.
- **Team tools:** Hive, Slite, Process Street, Paperform, Readwise, Raindrop.io, Avoma, Bubble, and Gelato.
### How it works
- **Documented from the source.** 4,801 operations, each written from the provider's official reference and checked against it by a separate review.
- **Checked on connect.** Each key is confirmed with one read-only call, so connecting never places a phone call, renders a file, or runs a scrape. Benzinga's check reads a single news headline, since every Benzinga endpoint is a licensed data product.
- **Your own server.** Flowise and Bubble take your app's address, and the key is only ever sent there, never to a private network address.
- **Regions and tenants.** Sauce Labs, Flatfile, DigiCert, InfluxDB Cloud, and Synthflow ask which region your account is in, and WATI uses your account's own API address.
See the [fifty integrations added before these](/changelog/2026-10-14-cursor-v0-harness-and-more).
## Cursor, v0, Harness, Neo4j, and 46 more integrations
> Fifty integrations joined the catalog, from Cursor, v0, and Harness to Neo4j, Dify, and Files.com, each with documented operations and a key check.
Oct 14, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-14-cursor-v0-harness-and-more
Fifty more integrations are live, each connected with an API key: developer platforms and coding agents, AI evaluation and vision, media and file processing, search and web data, email and messaging, and team tools. Find them in the [integrations catalog](/apps), including [Cursor](/apps/cursor), [v0](/apps/v0), and [Harness](/apps/harness).
### What was added
- **Developer tools:** v0, Cursor, Prisma, Convex, Trigger.dev, Neo4j, Sourcegraph, Harness, Honeybadger, ConfigCat, StatusCake, Builder.io, and Zeplin.
- **AI and agents:** Braintrust, Arize, Zep, Dify, Roboflow, Google Cloud Vision, Gladia, Luma AI, Wolfram|Alpha, Botpress, and Voiceflow.
- **Media and files:** imgix, GIPHY, remove.bg, TinyPNG, Shotstack, Transloadit, Files.com, and Brandfetch.
- **Search and web data:** Listen Notes, Zyte, Alpha Vantage, Foursquare, Geocodio, Moz, Bing Webmaster Tools, and Microsoft Clarity.
- **Email and messaging:** Benchmark Email, SMTP2GO, Woodpecker, LINE, and respond.io.
- **Team tools:** Fibery, Kintone, Everhour, Baremetrics, and Wufoo.
### How it works
- **Documented from the source.** 6,854 operations, each written from the provider's official reference and checked against it by a separate review.
- **Checked on connect.** Each key is confirmed with one read-only call. A few providers meter every call: Microsoft Clarity allows ten data exports a day per project, Foursquare's check is one Places search, and Alpha Vantage accepts any key on its free data, so its check only confirms a key was entered.
- **Your own server.** Sourcegraph and Dify take your instance's address, and the key is only ever sent there, never to a private network address.
- **Regions and tenants.** Honeybadger asks whether your account is in the US or EU, Shotstack whether the key is for its sandbox or production, and Files.com, Kintone, Fibery, and Wufoo use your own account address.
See the [fifty integrations added before these](/changelog/2026-10-13-rootly-jumpcloud-tomtom-and-more).
## Custom models: provider presets and a connection test
> Adding your own OpenAI-compatible model now starts from a provider preset, lists the models it serves, and tests the connection before members can pick it.
Oct 14, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-10-14-custom-model-endpoints
You can bring your own model to General Input from any endpoint that speaks the OpenAI Chat Completions API: OpenAI, Google Gemini, xAI, DeepSeek, Mistral, Groq, Together AI, Fireworks, OpenRouter, Cerebras, Moonshot, or your own deployment. Adding one in Settings, Models is now a guided form. Pick a provider and the URL fills in, choose the model from the list the endpoint serves, and test the connection before it reaches anyone's picker. Your provider bills you directly, and these models sit next to the hosted [models](/models) members already use.
### How it works
- **Provider presets.** Choosing a provider fills in its base URL and links to the page where you create an API key. Any other endpoint works with a URL you type.
- **Model list.** The form reads the models your endpoint offers, with each model's context window when the endpoint reports it, so you pick an ID rather than guess.
- **Connection test.** Before saving, we send the model one short message that asks it to call a tool, because agents work through tools. It reports how long that took, or what went wrong in plain words: a rejected key, an unknown model name, a model without tool calling, or a URL that isn't an OpenAI-compatible API. You can still save after a failed test.
- **Images and files reach the model.** A custom model marked as reading images or PDFs now receives them in chats and workflows. Before, that setting was saved but ignored.
- **Public endpoints only.** Our servers call your endpoint over the internet, so its URL must be publicly reachable.
### Also shipped
- Custom models get their own neutral icon instead of the General Input logo, which belongs to Auto.
- Each custom model row shows its model ID and host.
## Instantly triggers fire on replies, lead changes, campaigns
> All eleven Instantly triggers work again: new replies, lead status changes like Interested or Bounced, and completed campaigns now start your workflows.
Oct 14, 2026. Topics: Integrations, Workflows. https://www.generalinput.com/changelog/2026-10-14-instantly-triggers
Workflows started by [Instantly](/apps/instantly) triggers now run when the event happens. Eleven triggers were affected: New Reply Received, Campaign Completed, and the nine lead triggers, from Lead Marked Interested and Lead Booked a Meeting to Lead Bounced and Lead Unsubscribed. Most of them rarely or never fired, and fetching a sample while building a workflow always came back empty. A workflow you already built on one of them now runs with no change on your side.
### How it works
- **Lead triggers watch for status changes.** They ask Instantly for leads whose interest status changed in the last two days, so a lead imported weeks ago still fires the moment it is marked Interested.
- **Bounces and unsubscribes have room to arrive late.** Lead Bounced looks at leads contacted in the last 7 days and Lead Unsubscribed at the last 30, since a bounce or an unsubscribe click can land days after the email went out.
- **Each lead fires once per trigger.** A lead marked Meeting Booked starts that trigger's workflow one time.
- **Replies are read by arrival time.** New Reply Received looks at replies Instantly received in the last six hours and fires once for each, auto-replies included.
- **Every completed campaign counts.** Campaign Completed fires once for each campaign that reaches Completed, however long ago it was last edited. Trigger samples now return real leads, replies, and campaigns from your account.
### Also shipped
- Agents can now use Instantly's AI Sales Agent, AI Inbox Manager, AI Lead Finder, Deliverability Agent, saved search, and company list APIs, through 85 newly documented operations.
## iPhone conversations keep up through spotty service
> iPhone conversations now catch up on their own, retry through a weak signal, say when the network is the problem, and reconnect as soon as service returns.
Oct 14, 2026. Topics: Mobile. https://www.generalinput.com/changelog/2026-10-14-iphone-spotty-service
A conversation on the [iPhone app](/mobile) could open on the copy saved on your phone and, on a weak signal, miss a newer reply until you left and came back. Conversations now keep reading quietly until they are current, say so only when the network is in the way, and the app reconnects as soon as your phone has service again.
### How it works
- **Quiet unless something is wrong.** Opening a conversation reads in what it missed with no loading line. "Connecting…" or "Waiting for network" shows only after three seconds of trouble, the lists' words. One that can't load yet waits out the same three seconds and loads on its own once service returns.
- **Retries on its own.** A read lost on a weak signal tries again after one second, doubling up to thirty seconds, for as long as the conversation is open.
- **Back when service is.** When your iPhone reports its network is back, the live connection, your lists, and the open conversation catch up at once instead of waiting out their retry timers.
- **Quick actions are kept.** Marking a specialist read, pinning, and similar actions made while the connection drops are sent later instead of undone.
- **Replies arrive in one step.** When a run's end reaches the phone before its reply does, the typing bubble stays until the reply is in hand.
### Also shipped
- A reply that comes in while you have its conversation open on a bad connection is marked read, so it does not come back as unread.
- A second message or Stop in a new conversation, made before your first message got through, now stays in that conversation instead of starting another.
## JobDiva, iCIMS, Teamtailor, and 13 more recruiting tools
> Sixteen applicant tracking and staffing systems joined the catalog, from JobDiva and iCIMS to Teamtailor and Personio, each with a credential check.
Oct 14, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-14-jobdiva-icims-teamtailor-and-more
Sixteen recruiting and staffing systems are now in the integrations catalog. Specialists can search candidates and jobs, move people through pipelines, log notes and activity, and pull placements, timesheets and invoices into workflows and apps. Find them in the [integrations catalog](/apps), including [JobDiva](/apps/jobdiva), [iCIMS](/apps/icims), and [Teamtailor](/apps/teamtailor).
### What was added
- **Staffing and recruiting agencies:** JobDiva, Avionté, Ceipal, Vincere, Recruit CRM, Recruiterflow, Crelate, Loxo, and Manatal.
- **In-house hiring teams:** iCIMS, SmartRecruiters, Jobvite, Teamtailor, Pinpoint, and Zoho Recruit.
- **HR:** Personio, including its recruiting endpoints.
### How it works
- **Documented from the source.** 2,996 operations, each written from the provider's official API reference and checked against it by a separate review.
- **Checked on connect.** Each credential is confirmed with a sign-in or one read-only call. JobDiva, Personio, Ceipal, and Avionté trade your credentials for a short-lived token first, and a wrong password is sent only once.
- **Your account's address.** Teamtailor asks whether your account is in the EU or North America, Zoho Recruit uses your Zoho data center, and Pinpoint, Loxo, and Vincere use your own account address.
- **Provider-issued access.** iCIMS, Avionté, Jobvite, Ceipal, Recruiterflow, and Vincere issue API credentials on request rather than in their settings, and some plans gate the API: Manatal enables it by default on Enterprise Plus, Crelate and Recruit CRM on Business plans, and Personio on Core Pro.
See the [hundred integrations shipped alongside these](/changelog/2026-10-14-tableau-coinbase-kibana-and-more).
## The side panel shows everything a specialist can reach
> Beside a conversation, the side panel now lists the specialist’s files, apps, workflows, skills, and connections, with the conversation’s own files beneath.
Oct 14, 2026. Topics: Specialists. https://www.generalinput.com/changelog/2026-10-14-specialist-side-panel
The panel on the right of a conversation used to list only what that one conversation had touched: the resources it built or mentioned, the files that came and went, and the services it called. Every conversation now belongs to a specialist, so the panel shows the specialist instead. Its top card lists everything the specialist can see, ready to open without leaving the chat, and a second card beneath keeps the files that came into and out of this conversation.
### How it works
- **The specialist’s card on top.** Files, Apps, Workflows, Skills, and Connections, each with its count, the same things the specialist’s page holds.
- **Open in place.** A file, app, workflow, or skill opens in the side panel beside the conversation, [where chat files already open](/changelog/2026-10-13-chat-files-in-side-panel). A connection opens its own settings page.
- **The conversation’s card beneath.** Inputs are the files you attached and Outputs the files the specialist sent back. The card appears once the conversation has its first message.
- **There from the start.** A new conversation shows the specialist’s card before you type, so you can see what it can work with before you ask.
- **Room for both.** When a specialist holds a lot, the whole panel scrolls as one, with the conversation’s card below the specialist’s.
### Also shipped
- **Runs list their own files.** The panel beside a workflow or app run now shows only that run’s Inputs and Outputs.
## Tableau, Coinbase, Kibana, and 97 more integrations
> One hundred integrations joined the catalog, from Tableau, Coinbase, and Kibana to sevDesk and Gitea, each with documented operations and a key check.
Oct 14, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-14-tableau-coinbase-kibana-and-more
One hundred more integrations are live, each connected with an API key. Find them in the [integrations catalog](/apps), including [Tableau](/apps/tableau), [Coinbase](/apps/coinbase), and [Kibana](/apps/kibana).
### What was added
- **Developer tools:** Gitea, AppVeyor, Cloudsmith, BlazeMeter, GTmetrix, ilert, Instatus, Mezmo, NextDNS, MxToolbox, Xata, Pinata, Seam, UniFi, ntfy, Vast.ai, MailSlurp, Logo.dev, and IBM X-Force Exchange.
- **Data and AI:** Tableau, Kibana, DataRobot, Klipfolio, GoSquared, Knack, SeaTable, KoboToolbox, Softr, CrewAI, Eden AI, Wit.ai, PromptLayer, Linkup, Mathpix, NASA Open APIs, and API Ninjas.
- **Finance:** Coinbase, sevDesk, Lexware Office, Moneybird, Holded, Fixer, and Marketstack.
- **Sales and enrichment:** Pipeline CRM, CentralStationCRM, Simla.com, LeadIQ, Explorium, FullEnrich, Dropcontact, DaData, Better Proposals, and Tapfiliate.
- **Commerce and logistics:** BaseLinker, Loyverse, Lodgify, Rentman, Route4Me, OptimoRoute, GraphHopper, Geoapify, OpenCage, Jungle Scout, OpenSea, Givebutter, and PostGrid.
- **Marketing and messaging:** ContentStudio, Beamer, Userflow, Kontent.ai, Memberstack, Mighty Networks, UniSender, Yespo, MSG91, seven, Pushbullet, Chatwork, TinyURL, ClickMeeting, and Demio.
- **Documents and media:** ConvertAPI, Api2Pdf, Docparser, Copyleaks, Happy Scribe, SmugMug, Matterport, Penpot, Internet Archive, and Accredible.
- **Teams:** 7shifts, Breathe HR, TimeCamp, HackerRank, MOCO, Businessmap, HackMD, BugHerd, and Splitwise.
### How it works
- **Documented from the source.** 14,240 operations, each written from the provider's official reference and checked against it by a separate review.
- **Checked on connect.** Each key is confirmed with a call that sends, books, and converts nothing. Where a provider offers no free way to check a key, as with GraphHopper, Geoapify, and OpenCage, the check uses a single lookup from your quota.
- **Your own server.** Gitea, Kibana, Penpot, ntfy, SeaTable, KoboToolbox, DataRobot, and Tableau take your server's address, and the key is only ever sent there.
- **Regions and tenants.** Userflow and Knack ask where your account is hosted, and MOCO, CentralStationCRM, Simla.com, and Businessmap use your account's own address.
## Workspace database queries that hold up on real data
> Apps and workflows now look up long lists, search long phrases, and import wide tables without tripping database limits that small tests never hit.
Oct 14, 2026. Topics: Apps, Workflows. https://www.generalinput.com/changelog/2026-10-14-workspace-database-query-limits
Hosted [workspace databases](/changelog/2026-03-26-workspace-databases) refuse a query past a few fixed limits, and each of them hides during testing. An app that looked up one row per price-list line worked on a short test list and failed on the real one. Specialists now write queries that stay inside the limits however much data arrives, and when a query does go over, the error says how to fix it.
### How it works
- **Long lists go in as one value.** A lookup by hundreds of ids passes the whole list as a single JSON value instead of one placeholder per id, so it works past the limit of 100 values per statement.
- **Search that takes long phrases.** A search box matches with a substring check rather than a LIKE pattern, which is capped at 50 bytes and fails once someone searches for a long phrase.
- **Wide imports fit.** A table holds at most 100 columns, so a wide spreadsheet keeps its extra fields together in one JSON column.
- **Errors that name the fix.** A query refused for any of these limits, or for a statement over 100 KB, comes back with the limit it hit and the way around it, so the agent corrects its own query.
### Also shipped
- **Several statements, one transaction.** A query of several statements now runs all of them or none and answers with the last statement's rows, where before only the first statement's rows came back.
- **SQL Server text stays as written.** A dollar amount inside quoted text, such as 'Save $10', is no longer mistaken for a query parameter on connected SQL Server databases.
## Choose which models your workspace and Auto can use
> Settings, Models is now a live map of where every model comes from, with one switch per model that holds across hosted models, subscriptions, and custom ones.
Oct 14, 2026. Topics: Models, Collaboration. https://www.generalinput.com/changelog/2026-10-14-workspace-model-access
Settings, Models now shows your agent models as a live map of where they come from. Three sources sit on the left: models General Input hosts, members' Claude, ChatGPT, Gemini, Grok, and Muse subscriptions, and your custom endpoints. Every model sits in the middle with its switch, and lines run on into your workspace. Use it for cost control: keep a team on Auto and a few affordable [models](/models).
### How it works
- **One switch per model, across every source.** Turn a model off and nobody in the workspace can run it, whether from General Input, a member's subscription, or a custom endpoint. Another turns off General Input's hosted models entirely, keeping subscriptions and custom models.
- **See what's live.** A green line shows every way a model runs now; a model nothing serves says so.
- **Click a source to use it.** Subscriptions opens to connect your accounts, share them with the workspace, and turn your own models off. Custom opens to add or edit your endpoints' models.
- **Everyone sees it, admins change it.** Every member sees the map and can connect and share subscriptions; only admins switch models.
- **Auto follows your switches.** Auto tries connected subscriptions first, then hosted models, running only models that are on. A workflow set to a model that is off runs on Auto; a chat moves to the member's default or Auto.
### Also shipped
- Click a model for its card: its credits per 1K tokens, context window, and what it reads. Each provider lists its priciest first.
- The utility model gets its own section: pick the one small model that names chats and handles other background work.
- The "Default to subscription models" switch is gone: turn a model off to keep Auto off it.
## Agents can use your saved browser from their first step
> The browser is now a tool every agent starts with, so workflows with a saved browser no longer fail because the agent never set the browser up.
Oct 13, 2026. Topics: Agents, Workflows. https://www.generalinput.com/changelog/2026-10-13-browser-ready-from-first-step
We made the browser one of the tools every agent starts with, like running code or reading files. Until now, an agent had to load the browser guide before it could open a browser at all. Most runs did, but a scheduled workflow with a [saved browser](/changelog/2026-09-25-saved-browsers-power-button) would sometimes skip that step, decide it had no browser, and stop without opening a page, even though the browser and its saved logins were fine. That failure is gone: a workflow that follows a script now goes straight to opening the browser.
### How it works
- **Workflows with a saved browser** have the browser from their first step, whether they run on a schedule, from a trigger, or by hand.
- **Conversations with a [specialist](/templates/specialists)** always have the browser, for any site behind a login or without an API.
- **Workflows without a saved browser** don't see the browser at all, since they could never open one. To give a workflow a browser, attach one where you connect its other accounts.
- **The browser guide** is still there for agents to read before they drive an unfamiliar site by hand: keeping track of what was asked, checking each step, handling sign-in pages, and recovering from errors.
## Chat files open in the side panel beside the conversation
> On a wide enough screen, a file you click in a chat now opens in the right side panel, so you can read it and keep the conversation going.
Oct 13, 2026. Topics: Specialists, Agents. https://www.generalinput.com/changelog/2026-10-13-chat-files-in-side-panel
Clicking a file in a conversation with a specialist used to cover the whole screen. On a desktop window wide enough for the chat's side panel, it now opens there, where workflows, apps, and skills already preview, so you can read a report or check a spreadsheet while you keep talking.
### How it works
- **Every file in the chat.** Files you attach, files the specialist delivers, images in the transcript, and the Inputs and Outputs lists in the side panel all open in the panel.
- **A page made for a narrow panel.** Reports and markdown read edge to edge, Word documents reflow to the panel's width instead of shrinking, and PDF pages and images sit on a light canvas.
- **PowerPoint shows real slides.** Decks render with their layouts, colors, shapes, images, tables, and charts, where they used to show only their text.
- **Code reads like code.** Source files show line numbers and syntax colors and scroll sideways instead of wrapping, here and in the [full-screen file viewer](/changelog/2026-09-14-file-viewer). Spreadsheets right-align their figures.
- **A lighter full-screen viewer.** Full screen floats the file over a frosted view of the app instead of a black screen, with images, pages, and slides on soft shadows and documents on a white sheet.
- **Narrow windows keep full screen.** On a phone or a small window, files still open full screen.
### Also shipped
- **One viewer for every file.** A specialist's files, storage, memory uploads, and the files a workflow run used or produced now open in the viewer instead of downloading, and run files do in the iPhone app too. Download is still one click away.
## 50 new integrations: Braze, Telnyx, Airwallex, and more
> Fifty integrations joined the catalog, from Braze and Telnyx to Airwallex, Shortcut, and Fleetio, each with documented operations and a key check.
Oct 13, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-13-fifty-new-integrations
Fifty new integrations are live, each connected with an API key: marketing and messaging tools, CRMs, engineering platforms, field operations software, payments, and AI models. Agents and workflows call them like any other integration in the [catalog](/apps), starting with [Braze](/apps/braze), [Telnyx](/apps/telnyx), and [Airwallex](/apps/airwallex).
### What was added
- **Marketing and messaging:** Braze, Iterable, Omnisend, SendPulse, MailerSend, Mailtrap, Telnyx, Plivo, and Missive.
- **Sales and support:** Copper, Affinity, folk, Capsule CRM, Salesmate, Nutshell, Plain, Freshservice, and OnceHub.
- **Product and engineering:** Shortcut, Productboard, Postman, Statsig, Rollbar, BugSnag, Fullstory, Hex, Temporal Cloud, Sanity, Storyblok, Outline, Twenty, and Plausible.
- **Operations:** SafetyCulture, MaintainX, Fleetio, Connecteam, Toggl Track, and Teamwork.
- **Payments, data, and AI:** Paystack, Airwallex, Lemon Squeezy, TaxJar, Lob, Google Maps, DataForSEO, Cohere, OpenRouter, DeepSeek, xAI, and Recall.ai.
### How it works
- **Full API coverage.** 9,802 operations, each documented from the provider's official reference, so agents can reach endpoints beyond the common ones.
- **Checked on connect.** Saving a key runs one read-only call to confirm it works and records which account it belongs to. Keys with narrow permissions still connect.
- **Your own instance.** Outline, Twenty, and Plausible take your server's address, and the key is only ever sent there, never to a private network address.
- **Clearer errors.** When a connection fails, the message names the field to fix, or says when an address doesn't belong to the service.
## Gemini 4 Argon is available on launch day
> Google's Gemini 4 Argon is a hosted agent model with a 1M-token window, at introductory pricing of $2 input and $10 output per million tokens.
Oct 13, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-10-13-gemini-4-argon
Google released Gemini 4 Argon on October 13, and it's in General Input the same day. Every workspace can pick it for specialists, workflows, and chats, and it runs on our keys with no setup. If your Gemini subscription includes Argon, [Token Tunnel](/token-tunnel) offers it on the plan you already pay for. Compare it with every other model on the [models page](/models).
### How it works
- **Pricing.** Google's introductory rate runs through November 30: $2 input and $10 output per million tokens, and $0.10 per million cached input tokens. Google's list price doubles on December 1.
- **Window.** 1M tokens of context, with image, PDF, and video input.
- **Effort.** Argon defaults to high, and low and medium are there when you want a faster reply. In our tests on every route, high always reasoned before answering, while low and medium almost never did, even on hard problems.
- **Routes.** Google Vertex AI serves it first, with Vercel AI Gateway and OpenRouter behind it. When it's overloaded, runs fall back to Claude Opus 5.5, then GPT-6.1 Sol.
- **Token Tunnel.** Google offers Argon in Antigravity on AI Ultra and Enterprise plans. The tunnel lists it as soon as your signed-in Gemini account serves it, under the same name as the hosted model.
## Rootly, JumpCloud, TomTom, and 47 more integrations
> Fifty integrations joined the catalog, from Rootly and JumpCloud to TomTom, Svix, Runway, and Dub, each with documented operations and a key check.
Oct 13, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-13-rootly-jumpcloud-tomtom-and-more
Fifty more integrations are live, each connected with an API key: incident response and monitoring, webhooks and realtime messaging, AI models and speech, file and image processing, maps and weather, email marketing, and billing. Find them in the [integrations catalog](/apps), including [Rootly](/apps/rootly), [JumpCloud](/apps/jumpcloud), and [TomTom](/apps/tomtom).
### What was added
- **Developer tools:** bunny.net, Svix, Hookdeck, Ably, UptimeRobot, Pingdom, Rootly, Daily, Sendbird, and Lokalise.
- **AI and speech:** Weights & Biases, Writer, Scale AI, AI21 Labs, Kaggle, Runway, Rev AI, and Speechmatics.
- **Files and documents:** CloudConvert, Uploadcare, ImageKit, Bannerbear, and PDFMonkey.
- **Data and content:** Databox, Baserow, Grist, Prismic, Formbricks, Twelve Data, Semantic Scholar, and Serper.
- **Maps and weather:** TomTom, OpenWeather, Radar, and AccuWeather.
- **Marketing and sales:** Elastic Email, Moosend, EmailOctopus, Dub, Short.io, Mixmax, Salesflare, and Canny.
- **Operations and billing:** JumpCloud, Flutterwave, Maxio, Printify, Dovetail, Rocketlane, and Float.
### How it works
- **Documented from the source.** 5,620 operations, each written from the provider's official reference and checked against it by a separate review.
- **Checked on connect.** Each key is confirmed with one read-only call. Serper and Twelve Data have no free account lookup, so their check spends one credit on your plan.
- **Your own server.** Baserow, Grist, and Formbricks take your instance's address, and the key is only ever sent there, never to a private network address.
- **Regions and tenants.** JumpCloud, Speechmatics, and Maxio ask which region your account lives in, and Sendbird asks for your application ID, so every call goes to your own host.
See the [ten integrations added earlier today](/changelog/2026-10-13-unleash-crowdin-here-and-more).
## Unleash, Crowdin, HERE, Zulip, and six more integrations
> Ten integrations joined the catalog: Unleash, Crowdin, HERE, Zulip, Alpaca, ShipEngine, Persona, Langfuse, ngrok, and CoinMarketCap.
Oct 13, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-13-unleash-crowdin-here-and-more
Ten more integrations are live, each connected with an API key: feature flags, localization, maps and routing, team chat, trading, shipping, identity verification, LLM observability, tunnels, and crypto market data. Find them in the [integrations catalog](/apps), including [HERE](/apps/here) and [Crowdin](/apps/crowdin).
### What was added
- **Developer tools:** Unleash, ngrok, and Langfuse.
- **Content and communication:** Crowdin, including Crowdin Enterprise, and Zulip.
- **Location and logistics:** HERE (geocoding, routing, matrix and isoline routing, tour planning, traffic, and maps) and ShipEngine.
- **Finance and identity:** Alpaca (paper and live trading, plus market data), CoinMarketCap, and Persona.
### How it works
- **Documented from the source.** 2,007 operations, each written from the provider's official reference and kept to what an agent needs for the call.
- **Cloud or self-hosted.** Unleash, Zulip, and Langfuse take your instance or region address, and the key is only ever sent there, never to a private network address.
- **Checked on connect.** Each key is confirmed with one read-only call. HERE has no free account lookup, so its check is a single geocoding request on your plan.
- **Paper or live.** Alpaca asks which environment a key belongs to and uses the matching trading host.
See the [50 integrations added earlier today](/changelog/2026-10-13-fifty-new-integrations).
## Clear no-access messages, and admins can view any specialist
> Pages you can't open now say so right away, admins can view every specialist and skill, and every share dialog uses one clear layout.
Oct 12, 2026. Topics: Collaboration, Security. https://www.generalinput.com/changelog/2026-10-12-access-messages-and-admin-view
Opening another member's private specialist used to retry for about seven seconds, then show a raw "not found" error. Detail pages now tell the difference between something you can't open and something that doesn't exist, and they say so right away.
### How it works
- **Access versus missing.** Specialists, workflows, apps, skills, and connections show "You don't have access" at once, with no retries, when the item exists but isn't shared with you. Anything deleted, missing, or in another workspace says it isn't available. The dashboard and iPhone app use the same wording.
- **Admins can view every specialist and skill.** Workflows and apps already had an admin view; now Owners and Admins can open any specialist or skill in their workspace by link, as a viewer. Nothing is added to their own lists, and chatting in another member's specialist still needs a share, so the chat entry explains why it's disabled.
- **Links that work.** Specialist links in Analytics and on member profiles open the specialist's page.
### Also shipped
- One share dialog for everything (specialists, workflows, apps, skills, databases, storage, connections, and browsers) on the web and iPhone: invite people as chips, see who has access, switch between invited-only and everyone in the workspace, copy a link, or transfer ownership.
- Sharing now says "Private" instead of "Restricted."
## Decision API: classify and score items with probabilities
> A hosted decision service answers choice, score, and yes-or-no questions about any item with a probability for each answer, for workflows and apps.
Oct 12, 2026. Topics: Developers, Workflows. https://www.generalinput.com/changelog/2026-10-12-decision-service
Many automations hinge on a judgment call: which team should get this ticket, is this lead a good fit, does this email need a reply. The new decision service answers questions like these with a probability for every possible answer, so your workflow can act only when it's confident enough. Like every built-in service, it needs no API key of your own and bills to your workspace.
### How it works
- **Three kinds of question.** A choice among up to 255 options, a score on up to 10 ordered levels, or a yes or no. One request can ask several questions about the same item.
- **Any item.** Send text or JSON, up to 32,000 input tokens.
- **Thresholds you control.** A new decisions skill teaches specialists how to phrase questions and set a confidence threshold for each action.
- **Stable models.** A request always names its model and is never silently moved to a newer one, because thresholds are tuned per model. At launch the model is TypeSafe's Jev, and you can bring your own TypeSafe key.
- **Clear failures.** Errors come back as one of six codes, and a provider's raw error never reaches you.
- **Proven in-house.** Our own Auto model routing and New task routing run on the same service.
### Also shipped
- Calls to our chat completion service can now name an exact model, such as Claude Haiku 5.5, Sonnet 5.5, Opus 5.5, or Gemini 3.8 Flash.
## Meta Muse Wonder arrives, and Muse Code joins Token Tunnel
> Meta's Muse Wonder is available as a hosted agent model with a 1M-token window, and Muse Code is the fifth subscription Token Tunnel connects.
Oct 12, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-10-12-muse-wonder-and-muse-code
Meta released Muse Wonder on October 12, and it's in General Input the same day. You can also connect a Muse Code subscription through [Token Tunnel](/token-tunnel), alongside Claude, ChatGPT, Gemini, and Grok, and run Muse models on the plan you already pay for. Read about [Meta's earlier open-source Muse release](/blog/meta-open-sources-muse-spark-1-2).
### Muse Wonder
- A 1M-token context window, $4 input and $20 output per million tokens, $0.20 per million cached input tokens, and effort from minimal to xhigh.
- A default agent model, so specialists can run on it.
- Served through Vercel AI Gateway and OpenRouter. When it's overloaded, runs fall back to Claude Opus 5.5, then GPT-6.1 Sol, both in the same price tier.
### Muse Code on Token Tunnel
- Sign in with a device code from Meta, the same flow Grok uses.
- The tunnel offers Muse Wonder and Muse Spark 1.3. Muse Code's contributor variants, which train on your data, are never offered, and a test enforces it.
- Muse follows instructions reliably only when they arrive as a real system prompt (5 in 5 runs, versus 1 in 5 otherwise), so the tunnel sends them that way. Our tunnel test suite caught this before release.
### Also shipped
- Claude Haiku 5.5 on a Claude subscription now reliably answers as your specialist, held 10 of 10 times in testing.
## Real-time sync across web and iPhone with one sync engine
> Every list on the dashboard and iPhone now draws from one sync engine: your actions show instantly, others' changes arrive in about two seconds.
Oct 12, 2026. Topics: Platform, Mobile. https://www.generalinput.com/changelog/2026-10-12-one-sync-engine-for-every-list
The dashboard and the [iPhone app](/mobile) held the same lists (specialists, Home, Activity, conversations, resource cards, pins, badges) in about a dozen separate caches, and every send, stop, read, and live event had to remember to update each one. When one was missed, a specialist looked idle while it worked, or a finished run still looked busy. Both apps now draw every list from one sync engine per workspace, kept in step by a server change feed.
### How it works
- **A numbered change feed.** For each member, the server rebuilds what their lists show, records what changed under a sequence number, and clients pull everything past the number they have, with live updates in between. It follows Replicache's client-view approach.
- **Your actions draw at once.** Sends, stops, reads, pins, renames, and moves appear immediately and stay put until the server confirms them; only a refusal undoes one.
- **Opens already drawn.** The iPhone keeps the engine on the device and each browser tab keeps a copy locally, so the app draws before any network response.
- **Force-quit included.** When a push arrives, the iPhone fetches changes in the background, so even a force-quit app opens with the right run status.
- **Shared resources follow.** A change to a shared workflow or specialist reaches everyone whose lists show it, within about two seconds.
- **Tested hard.** A lifecycle harness runs 300 randomized sessions against a fake iPhone and a network that delivers answers in any order.
### Also shipped
- Send videos from your iPhone, up to 512 MB.
## Transcription API: turn audio and video into text
> A hosted transcription service turns audio and video files up to 1 GB into text, with speaker labels, word timings, and no charge for failed jobs.
Oct 12, 2026. Topics: Developers, Workflows. https://www.generalinput.com/changelog/2026-10-12-transcribe-service
Workflows and apps can now turn recordings into text with a built-in transcription service: meeting recordings, voicemails, interviews, podcast episodes, or any audio or video file. Like every built-in service, it needs no account of your own and bills to your workspace.
### How it works
- **Start a job.** Send a link to the audio, or upload the file itself (audio or video, up to 1 GB), and you get back a job ID.
- **Get the transcript.** Check the job for its status, full text, detected language, and audio length, plus speaker turns when speaker labels are on and word-level timings when you ask for them.
- **Better accuracy on your vocabulary.** Pass key terms, such as product names or industry jargon, to bias recognition: up to 1,000 on the most accurate model and 200 on the other.
- **Two models.** Choose between two AssemblyAI models, with the provider kept behind a seam so more can be added.
- **Fair billing.** The price is locked when a job starts, each job is billed once, and a failed job is free.
Use it from code workflows, app handlers, and sandbox scripts, and pair it with the [decision service](/changelog/2026-10-12-decision-service) to route calls by what was said.
## Agent Firewall blocks now say who can lift them and where
> When the Agent Firewall blocks an action, the agent now names the rule, links the setting that allows it, and says exactly who can change it.
Oct 11, 2026. Topics: Security. https://www.generalinput.com/changelog/2026-10-11-agent-firewall-block-messages
When the [Agent Firewall](/agent-firewall) blocked a request, the agent used to be able to say only "ask the workspace owner," which left people guessing. Every firewall refusal now explains itself: which firewall stopped the action, where it can be allowed, and who is allowed to change it.
### How it works
- **One shared answer.** For each firewall that denied an operation (the workspace's, the connection's own, or both), the refusal names the page that allows it: Settings, Agent Firewall for workspace rules, or the connection's own page.
- **Who can lift it.** The refusal says whether you can change the rule yourself, whether it takes a workspace Owner or Admin, or names the connection's owner. Only a connection's owner can change that connection's firewall; a workspace admin can't.
- **Everywhere agents run.** The same answer comes from the sandbox's outbound proxy, the `geni exec` command, and the integration relay, and the agent relays it with a link.
- **Warned before trying.** When an agent looks up an API operation that's blocked for a connection, the docs say so before it makes the call.
- **Never fails open.** Older sandboxes ignore the new details, and an unreadable policy falls back safely, so a version mismatch can't switch enforcement off.
## Zero-downtime app updates and faster cold starts
> An open app keeps running while a new build lands, sessions survive long calls, and a cold app's first request answers without polling delays.
Oct 11, 2026. Topics: Apps, Platform. https://www.generalinput.com/changelog/2026-10-11-apps-stay-up-through-publishes
When someone published a new version of an app you had open, the dashboard unmounted it for every build, failed builds included, and the iPhone app could run old screens against new server code for up to a minute. Viewers refreshing at once could also restart the app's server in parallel and disrupt each other. We rebuilt the session model under apps.
### How it works
- **The served build.** Each app records which build its viewers get, and that pointer moves only when a build succeeds. On the dashboard and iPhone alike, the running app stays up until the next build's server is ready, then reloads on the same page. A failed build never takes down the version you're using.
- **One restart.** After a publish, the first caller claims the server restart and everyone else waits for it.
- **Session identity.** Every server session gets a new ID that guards its writes, so a stale caller can't overwrite its replacement, and an unusable session is replaced in place.
- **Leases cover the longest call.** Calls in flight are counted, so a long export is no longer cut off when a short call ends.
- **Faster cold starts.** The app server starts and passes its health check in one step, and a first call that lands while it boots waits up to 10 seconds and returns directly.
### Also shipped
- App JavaScript and CSS are compressed, 68 to 78% smaller: a minimal app's first-paint bundle dropped from about 283 KB to 85 KB.
## Reliable conversation compaction for long AI chats
> Compaction now always lands a checkpoint, falling back to another model or a labeled excerpt, so one failed summary can't break a conversation.
Oct 11, 2026. Topics: Agents, Workflows. https://www.generalinput.com/changelog/2026-10-11-compaction-always-checkpoints
When a conversation outgrows its model's context, a summarizer writes a checkpoint and the conversation continues from there. In production, the default utility model could spend its whole output budget on hidden reasoning and stop before the summary was complete. Compaction then failed the run, and every later message in that conversation failed the same way. Compaction now always lands a checkpoint, and never writes a partial one.
### How it works
- **Bounded summaries.** The summarizer runs at its lowest thinking setting with a 16K-token output cap and a prompt asking for under 4,000 words.
- **A fallback chain.** If the utility model's summary errors, is cut off, or comes back empty, the run's own model writes it instead, and a context overflow retries once with half the input.
- **A last resort.** If no summarizer finishes, the checkpoint is a word-for-word excerpt of the start of the conversation and the latest work, clearly labeled, so the conversation keeps working.
- **Clean interruptions.** Cancel, timeout, and handoff abort the summary without writing anything, and Stop now interrupts a compaction in progress.
### Also shipped
- **Long workflows keep platform access.** A code workflow's platform key now lasts as long as the run itself, instead of expiring after 10 minutes.
- **Viewing a workflow never turns it off.** Opening a workflow's page mid-edit could disable it; reads no longer change anything.
- Fixes from our latest build eval: 22 builds on Claude Sonnet 5.5, all goals met.
## Phone trackpad control for live browsers, plus a watch mode
> Phone takeovers gain pinch zoom, tap to click, drag, right-click, and a clipboard, and every live browser gets real browser controls and a watch mode.
Oct 11, 2026. Topics: Agents, Mobile. https://www.generalinput.com/changelog/2026-10-11-live-browser-trackpad-and-chrome
Sign-in cards and AI connection sign-ins hand you a live browser so you can log in to a site yourself. On a phone, text sent into that page was typed key by key, which was slow and failed past 1,024 characters. The phone is now a trackpad over a zoomable page, and every live browser, on any screen, is framed in the same browser controls.
### Phone takeover
- Pinch to zoom up to 5x the fitted page, and pan or scroll with two fingers.
- Tap clicks and a quick second tap double-clicks. Tap then drag holds the button down for selecting text, sliders, and maps. A two-finger tap or long hold right-clicks.
- A clipboard menu copies the page's selection to your phone or pastes text from it in one step.
- A help sheet explains every gesture.
### Browser controls everywhere
Tabs, back, forward, reload, and the address bar frame the page wherever a browser appears, at one of three levels of access:
- **Drive:** your own browser, fully usable.
- **Drive in place:** an AI connection's sign-in, where the address is visible but can't be edited.
- **Watch:** an agent's browser inside a conversation or run. Everything is visible and nothing is pressable.
Watch mode is new: anyone who can read the conversation can see the agent's browser, and watching never affects the agent's session.
## Claude Haiku 5.5 becomes the default fast model
> Claude Haiku 5.5 replaces Haiku 4.5 with a 1M-token window and an effort setting, becomes a default agent model, and leads Auto's Fast tier.
Oct 8, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-10-08-claude-haiku-5-5
Anthropic released Claude Haiku 5.5 on October 7, and it reached General Input on October 8. It replaces Haiku 4.5 everywhere, and it's the first Haiku to be a default agent model rather than opt-in, so specialists can run on it. It appears in the model picker, the hosted API catalog, and the Subscription group when you connect a Claude plan through [Token Tunnel](/token-tunnel). See the [models page](/models).
### How it works
- **The model.** A 1M-token context window (Haiku 4.5 had 200K) and adaptive thinking with an effort setting that defaults to medium. Anthropic lists it at $0.10 input and $0.50 output per million tokens for prompts up to 100K tokens, and five times that above.
- **Saved selections move over.** Any workflow, specialist, or setting that named Haiku 4.5 now runs on Haiku 5.5.
- **Auto, hosted.** The Fast tier now tries Claude Haiku 5.5 first, then GPT-6 Luna, Gemini 3.8 Flash, and Grok 4.7.
- **Auto on your subscriptions.** When Auto can use a model you've connected through Token Tunnel, each tier now runs one class above its hosted counterpart, because subscription usage isn't billed per token: Fast uses Claude Sonnet 5.5 or GPT-6.1 Sol, Balanced uses Opus 5.5 or GPT-6 Astra.
### Also shipped
- A connected device that only serves Haiku 4.5 is no longer presented as Haiku 5.5.
## Lazy-loading agent workspaces with a real filesystem
> A cold sandbox now restores only file metadata and fetches file contents as the agent reads them, so large workspaces no longer delay the first command.
Oct 8, 2026. Topics: Agents, Platform. https://www.generalinput.com/changelog/2026-10-08-lazy-workspace-filesystem
Every conversation gives its specialist a persistent work folder in its sandbox. Until now, a cold sandbox fetched that folder's contents before the agent could act, so the bigger the workspace, the longer the wait. As of October 8, the work folder is a real filesystem, built on open-source JuiceFS, that restores only names and metadata when the sandbox starts and fetches each file's data the first time something reads it.
### How it works
- **Lazy restore.** Metadata is restored with the sandbox, file data is fetched from storage on first access and cached, and small files are prefetched in the background so a full scan doesn't pay a round trip per file.
- **As fast as a native sandbox.** A restore is ready in about the time a fresh sandbox takes to boot, for workspaces up to 5,000 files and 128 MiB.
- **Real filesystem behavior.** Large files, random reads, memory mapping, permissions, and links behave as on a local disk, so shell and file tools need no changes.
- **Safe checkpoints.** Checkpoints run periodically and when a run pauses, completes, or hands off; a failed checkpoint keeps the previous one and retries.
- **Isolation.** A sandbox never holds our storage keys. It gets credentials for its own workspace only, valid for at most 24 hours, and a storage gateway checks every request.
- **Limits.** A workspace holds roughly 190,000 files; past that, the agent is told to remove rebuildable folders like node_modules.
## New task: describe the work and General Input routes it
> Start with the task, not a specialist. Your specialists are ranked against what you need, and the work goes to the best fit or a new specialist.
Oct 8, 2026. Topics: Specialists. https://www.generalinput.com/changelog/2026-10-08-new-task-routing
Starting a conversation used to mean choosing a specialist first. Now you start with the work. New task, in the sidebar, the Home menu, and the specialist list, opens the full chat composer. When you send, General Input ranks the specialists you can access and asks where the task should go: pick the recommended specialist, another one, or a new specialist that's created, named, and described from your message and starts on the task right away. The same flow ships on iPhone.
### How it works
- **Ranking.** A decision model (Jev, from TypeSafe) classifies your message against up to 40 of your most recently active specialists plus a "needs a new specialist" option, weighting each specialist's name as the main signal of its scope.
- **Nothing hidden.** Most tasks get one featured pick, but every specialist with a reasonable chance is listed too, and Browse all specialists is always there.
- **Never blocking.** Ranking has a 2.5-second budget; if it's slow or unavailable, you see your recent specialists instead.
- **Onboarding.** New workspaces end onboarding on the same composer: "What do you want to tackle first?"
### Also shipped
- Specialists can [read your other conversations](/changelog/2026-10-08-read-other-conversations) when it helps.
- Redesigned specialist pages, and a sidebar that shows only the specialists worth a glance.
## Specialists can read your other conversations for context
> A specialist can now look up what you've discussed with your other specialists, reading only your own chats, so you don't have to repeat yourself.
Oct 8, 2026. Topics: Specialists, Agents. https://www.generalinput.com/changelog/2026-10-08-read-other-conversations
You shouldn't have to repeat yourself to every specialist. Since October 8, a specialist can check what you've been discussing elsewhere: if you worked out a plan with your Chief of Staff on Monday, the specialist that carries it out on Tuesday can read that conversation instead of asking you again.
### How it works
- **Two tools.** One lists your conversations across every specialist, newest first, optionally narrowed to one specialist or filtered by title. The other loads up to 25 of them into the specialist's workspace.
- **Read on disk, not in context.** Loaded conversations are written as files the specialist can search, so a long thread doesn't flood the model's context.
- **Your chats only.** A specialist reads only the conversations of the person it's talking to, and only in runs that belong to that person, never in a background run acting for someone else.
- **Clean transcripts.** Loaded conversations include only what you and the specialist said, each line labeled with its real speaker and time, without tool calls or hidden reasoning.
- **Reference, never instructions.** Old conversations are marked as reference material, so text quoted in an earlier chat is never treated as a new instruction.
Pairs with [New task](/changelog/2026-10-08-new-task-routing), which routes each new request to the right specialist.
## Credit spend analytics by specialist, resource, and member
> Analytics is now one spend view in credits: group by specialist, resource, activity, segment, or member, compare periods, and read every receipt.
Oct 7, 2026. Topics: Collaboration. https://www.generalinput.com/changelog/2026-10-07-analytics-spend-view
Workspace owners ask Analytics two questions: what are we spending credits on, and who is spending them? The old page spread the answers across four tabs. We rebuilt it as a single spend view measured in credits, with runtime and run counts beside spend, since spend alone doesn't measure productivity. An Overview and a Logs tab replace the old Members, Executions, Security, and Usage tabs.
### How it works
- **Toolbar.** A member filter and a rolling window from 24 hours to 90 days. The window, grouping, and filter live in the URL, so a link opens the same view.
- **Tiles.** Credits spent, executions, and AI runtime for the window, each compared with the previous window of the same length.
- **Group by.** Specialist, Resource, Activity, Segment, or Member, as a chart over time and a ranked list. Resource ranks individual workflows, apps, and skills, with a workflow's runs, builds, and upkeep collapsed into one row. Activity groups spend by kind of work and covers 100% of it.
- **Credits or usage.** Switch the unit from credits to AI runtime or operation counts, so an expensive member reads differently from an unproductive one.
- **Logs.** Itemized receipts (what, who, when, credits) right in the page.
### Access
Owners and Admins see the whole workspace; every other member sees only their own spend, enforced on the server for every endpoint behind the page.
## Offline-ready iPhone app with on-device conversations
> Chat lists and recent conversations are cached on your iPhone, notifications fetch the reply before you tap, and reopening loads only what changed.
Oct 6, 2026. Topics: Mobile, Specialists. https://www.generalinput.com/changelog/2026-10-06-mobile-device-cache
The [iPhone app](/mobile) now opens on what it already has. Chat lists and recent conversations live on the device, a notification fetches its conversation as it arrives, and coming back reads only the messages you missed. Tapping a reply notification opens on the reply, even in airplane mode.
### How it works
- **Device cache.** A local SQLite store holds your chat lists, recent conversations, and signed-in identity. Twenty conversations are restored at launch and up to 200 are kept; the rest load when opened. The cache is scoped per server, user, and workspace, and wiped on sign-out.
- **Delta reads.** Reopening a conversation fetches only messages after the last one you have, or the newest page if a lot is new.
- **Notification extension.** When a push arrives, a Notification Service Extension downloads the conversation, so the reply is already on the phone when you tap.
- **Away-time messages appear quietly.** Only messages you watch arrive live animate in; anything that landed while you were away is simply there.
- **No stale "working."** Typing indicators and Stop appear only after a fresh check confirms the run is still going.
### Also shipped
- Messages you send while a specialist works wait in a tray above the composer, where you can edit or remove them.
- Word documents preview as laid-out pages, rendered in an isolated frame.
- Links open full screen in the native Safari view.
## Accurate OAuth connection results instead of popup guesses
> The dashboard now asks the server how each connect attempt ended, so failed or abandoned connections no longer count as success.
Oct 6, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-10-06-oauth-connect-outcomes
Connecting an integration from the dashboard, or from a connect card in chat, now reports what actually happened. The server records how every connect attempt ended, and the dashboard waits for that answer instead of inferring it from the sign-in popup.
Before this change, the dashboard learned the result through a message from the popup. The browser security policy both the dashboard and the API send cuts that link, so the message often never arrived, and the fallback treated a closed-looking popup as success. In practice, failures and abandoned popups counted as connected, and a chat connect card could resume a specialist with no credential at all.
### How it works
- **One attempt, one record.** Starting a connection returns an attempt ID alongside the sign-in link, for every kind of OAuth connection.
- **A status check.** The dashboard checks the attempt every second until it's completed (with the new credential), failed (with its error), or expired after 10 minutes.
- **Owner-only.** Only the member who started an attempt can read it.
- **Security headers unchanged.** We kept the strict browser policy and moved the source of truth to the server.
### Also shipped
- New integrations: DocuWare, with a new-document trigger, and TalentLynk, with a campaign-status trigger.
## New workspaces start with $30 in free credits
> The welcome grant for a new General Input workspace rises from 800 credits ($5) to 4,800 credits ($30), a free month of one Pro seat.
Oct 6, 2026. Topics: Pricing. https://www.generalinput.com/changelog/2026-10-06-welcome-credits-30
Every new workspace now starts with 4,800 credits, worth $30, instead of 800 credits ($5). That's exactly one Pro seat's monthly allowance, so it amounts to a free first month of one Pro seat: enough to connect your tools, build your first specialists, and run real workflows before paying anything. See what a seat includes on the [pricing page](/pricing).
### How it works
- **One source for the number.** The grant is defined once, at our fixed rate of 160 credits per dollar, and the onboarding grant, the welcome cards on the dashboard and iPhone, and the pricing page's free tier ("4,800 free credits to start") all read that value, so they can't disagree.
- **Specialists know too.** The built-in help skill specialists use to answer product questions describes the new grant.
- **Who gets it.** Workspaces that accept their welcome credits after this release receive 4,800. Workspaces that already received 800 keep their existing balance.
- **Unchanged.** The low-balance warning still triggers at 160 credits.
New to General Input? Follow the [crash course](/blog/getting-started-crash-course) to sign up, connect your tools, and build your first workflow and app.
## More reliable browser agents, measured on WebArena-Verified
> The browser agent acts on stable element references and verifies batched form edits, and every change is scored on 468 WebArena-Verified tasks.
Oct 4, 2026. Topics: Agents. https://www.generalinput.com/changelog/2026-10-04-browser-agent-webarena
We reworked how the browser agent acts on pages and how long browsing runs hold their context, and we built a harness that scores browser changes on a public benchmark. The changes target three failure modes: forms left half-filled, observations that go stale after navigation, and long runs that lose their prompt cache.
### How it works
- **Stable element references.** The agent acts on the exact element it saw instead of re-finding it by selector.
- **Checked batches.** Several form edits apply in one guarded batch: each value is read back, a page that changed underneath is detected, and a batch that stops partway reports exactly what landed.
- **Settled observations.** After each step, the runtime waits for navigation and transitions to settle, and on large pages it puts the foreground content first.
- **Cache-friendly runs.** History within a run only grows, which keeps the prompt cache valid. An earlier version that pruned old screenshots cut cache reads from 90% to 65%, so we dropped it.
### How we measure it
The harness runs WebArena-Verified through our real browser stack: 468 of the 812 official tasks across Reddit, Shopping Admin, and GitLab clones, each in a fresh container with pinned versions. An October 4 sweep with Claude Sonnet 5.5 at high effort passed 349 of 468 (74.6%), the level future changes must beat.
### Also shipped
- Specialists can read workflow run history at any time, without loading a building skill.
## Instant chat updates and durable message sends on iPhone
> Sending, starting, stopping, and reading now update every list instantly, and iPhone sends retry for up to a day without ever duplicating.
Oct 4, 2026. Topics: Specialists, Mobile. https://www.generalinput.com/changelog/2026-10-04-instant-chat-durable-sends
Sending a message, starting a conversation, stopping a run, and reading a reply now update the sidebar, conversation lists, and Home immediately on the dashboard and in the [iPhone app](/mobile), instead of waiting for a round trip to the server. On mobile, a message you send keeps trying for up to a day, through dead zones, backgrounding, and relaunches, and a retry never creates a second copy.
### How it works
- **Same rows on both sides.** The code that builds conversation rows on the server is shared with the apps, so a row drawn optimistically matches the one the server returns. A failed action rolls back only what nothing newer has replaced.
- **A persisted outbox.** Every mobile send is written to the device first, then retried with backoff on network errors, and immediately when the app returns to the foreground. Messages in one conversation go in order.
- **Deduplicated on the server.** Each send carries a request ID, and the server keeps its receipt for a day, so a retry of a send that already landed gets the original result back.
- **Preloading.** The app fetches a conversation before you open it: when its run finishes, when a push arrives, and as soon as you tap.
### Also shipped
- Notifications clear on your phone when you read the item on another device, and badges stay in step.
- Conversation rows show what each one is doing, and specialist rows count the conversations that need you.
## Share a Token Tunnel subscription with chosen workspaces
> Owners of a hosted Token Tunnel account can share it with selected workspaces, so members there run chats and workflows on it. The login stays private.
Oct 2, 2026. Topics: Models, Collaboration. https://www.generalinput.com/changelog/2026-10-02-share-token-tunnel-accounts
A hosted [Token Tunnel](/token-tunnel) account (a Claude, ChatGPT, Gemini, or Grok subscription you signed in to through General Input) used to serve only its owner. You can now share each account with specific workspaces. Every active member of a workspace you add can use that subscription in chats and workflows, on the same plan limits, and you can remove a workspace at any time.
### How it works
- **Per-account sharing.** In Settings, under AI connections, each provider account has a Share action with a workspace picker. Recipients see the provider, its status, and your display name, never the provider account's email.
- **One model pool.** Your own accounts and accounts shared into the workspace form a single pool; requests spread across it, and a rate limit or connection failure falls through to another eligible account.
- **Pinned routing.** A shared request goes only to that one account, so retries never spill onto the owner's private accounts.
- **Live permission checks.** Both members and the share itself are checked before every request, so deactivating either one takes effect on the next attempt.
- **Clean revocation.** Disconnecting an account removes its shares, and reconnecting doesn't restore them.
### Also shipped
- GPT-6.1 Sol replaced GPT-6 Sol everywhere Sol is used, with saved selections moved automatically.
- Long runs on subscription models no longer fail while condensing the conversation, because the summary now streams.
## Faster agent workspaces: content-addressed file storage
> A conversation's files are now stored as deduplicated, content-addressed objects, so large workspaces restore quickly instead of re-downloading one archive.
Oct 1, 2026. Topics: Platform, Agents. https://www.generalinput.com/changelog/2026-10-01-content-addressed-workspaces
Every conversation with a specialist has a workspace: the files it keeps between turns. We used to save that folder as a single archive, so every cold start downloaded, unpacked, and later re-uploaded the whole thing. That scaled badly: one design conversation with 175 files and 260 MB of images saw its sandbox start climb from 2 to 15 seconds to more than 200 before it failed outright. Workspaces are now stored the way Git stores a repository, as content-addressed objects plus a manifest, synced incrementally.
### How it works
- **Stored once.** Each file is stored once by its content hash, and small files are grouped into packs, so a project with thousands of tiny files doesn't become thousands of requests. Unchanged files are never uploaded twice.
- **No lost work.** A new manifest is published only if the previous one is still current, so a late or stale save can't overwrite newer work or bring back deleted files.
- **Restores straight to the sandbox.** Files over 25 MiB upload from the sandbox itself instead of passing through our servers.
- **Migrated automatically.** Older workspaces convert the next time they're used.
### Also shipped
- When a workspace grows past 25 MiB in one file or 128 MiB in total, the agent sees the largest files and is guided to archive inactive ones to storage you've connected.
- Vercel joined E2B as a supported sandbox provider for deployments that configure it.
## Reliable image uploads in chat: no overwrites or file cap
> Same-named photo uploads no longer collide, the 20-file attachment cap is gone, images are resized for AI models, and oversized requests compact.
Sep 30, 2026. Topics: Agents. https://www.generalinput.com/changelog/2026-09-30-image-heavy-chats
Image-heavy conversations could go wrong four separate ways. Pasted phone photos could replace or corrupt each other, a large folder hit a fixed attachment cap, a request with many big images was rejected outright, and a request over a size limit ended the run. Between September 26 and 30 we fixed each one where it happened.
### How it works
- **Every upload gets its own path.** Phone photos are all called image.png, and paths came from the filename alone, so small uploads replaced each other and two large ones could write into the same file, leaving the agent with an image that was black below its top fifth. Your preview was always right; only the agent's copy was damaged. Each upload now lands in its own folder derived from its storage key.
- **No 20-file cap.** Attachments are planned against separate budgets for disk and for what's held inline, so a large folder works without flooding storage.
- **Images resized before sending.** Anthropic rejects a whole request when it carries more than 20 images and any is over 2,000 pixels on a side, so every outbound image is scaled to at most 2,000 pixels. For Claude this costs no quality.
- **Too big means compact, not fail.** A request over a size limit, such as 18 screenshots in one turn, now condenses the conversation and retries.
### Also shipped
- Condensed conversations keep your latest message word for word.
## Sign in with Apple on the web and iPhone
> Sign up, sign in, and accept invitations with your Apple ID, next to Google, and link or unlink Apple from your account settings.
Sep 30, 2026. Topics: Security, Mobile. https://www.generalinput.com/changelog/2026-09-30-sign-in-with-apple
You can now sign up, sign in, and accept invitations to a workspace with your Apple ID, alongside Google. Existing members can link Apple to their account, or unlink it, from account settings. The [iPhone app](/mobile) uses the same browser sign-in and device approval it already had.
### How it works
- **Fresh credentials every time.** The server signs a new, short-lived Apple client secret for every token exchange and refresh, so a long-running server never presents an expired one.
- **Strict origins.** Apple's sign-in callback is accepted only on its dedicated endpoint.
- **Clean unlinking.** Apple tokens are revoked before you unlink Apple or delete your account. If Apple fails temporarily, nothing is removed and you can retry.
### Also shipped
- Copying a pending invite's link always creates a fresh seven-day link.
- Google sign-in shows its account chooser, so you can switch Google accounts.
- Mobile sign-in stays open through brief network failures instead of closing midway through a password or two-factor prompt.
## General Input for iPhone, now in public beta on TestFlight
> Chat with your specialists, approve runs, and get push notifications on iPhone. The native app is open to everyone through a public TestFlight beta.
Sep 29, 2026. Topics: Mobile, Specialists. https://www.generalinput.com/changelog/2026-09-29-iphone-app-public-beta
General Input is on iPhone. The native app, open to everyone through a public TestFlight beta linked from the [mobile page](/mobile), puts your specialists in your pocket: chat with them, answer their questions, approve paused runs, and check on your workflows from anywhere.
### What's in the app
- **Home and Activity.** Your pinned work, conversations, and anything waiting on you, matching the dashboard.
- **Native chat.** Streaming replies, attachments, model selection, voice dictation, @-mentions, and answer cards for an agent's questions and approvals.
- **Your specialists and resources.** Browse specialists and their workflows, apps, and skills; open apps full screen; and launch, cancel, continue, or retry runs.
- **Push notifications** for runs that finish, fail, or need your input.
- **Easy sign-in.** Sign in through your browser with a quick device approval.
- **Built for iOS.** Native tabs with Liquid Glass on iOS 26, plus native menus and sheets.
The app keeps improving every week: since launch it gained [offline conversations](/changelog/2026-10-06-mobile-device-cache) and [real-time sync with the web](/changelog/2026-10-12-one-sync-engine-for-every-list).
## Introducing specialists: AI workers that each own one job
> Projects are now specialists: AI workers that own one job and build and run its workflows, apps, and skills. Every member gets a Chief of Staff.
Sep 27, 2026. Topics: Specialists, Collaboration. https://www.generalinput.com/changelog/2026-09-27-introducing-specialists
General Input now runs on specialists. A specialist is an AI worker that owns one job in your business, such as weekly reporting or lead follow-up. You describe the outcome in your own words, it asks what it needs to know, and it builds and runs the skills, workflows, and apps the job takes. Projects became specialists and kept everything they held. Read [why we made the change](/blog/introducing-ai-specialists).
### The model
- **A Chief of Staff for every member.** It answers questions across every connected tool, knows what each specialist owns, handles one-off work, and sets up new specialists for anything recurring.
- **Every resource has one home.** Each workflow, app, skill, file, and conversation lives in exactly one specialist.
- **One page per specialist,** holding its conversations, workflows, apps, skills, files, memory, and connected accounts.
- **Memory and files.** Each specialist keeps a plain-language memory of up to 12,000 characters for its rules and methods, and it never overwrites an edit it hasn't seen.
- **Members and privacy.** Sharing a specialist shares its resources, memory, and files; each member's conversations stay private.
### Starting a specialist
Start from a setup conversation, or from one of 65 [specialist templates](/templates/specialists), each built around a single job with its memory, sample files, skills, workflows, apps, and tests. Twenty-eight are workstream specialists for specific industries.
Nothing was deleted or turned off: each project became a specialist with the same contents, and loose work moved to each member's Chief of Staff.
### Also shipped
- A Chat command palette in the sidebar: pick a specialist, then one of its conversations.
## Self-serve account deletion on web and iPhone
> Delete your General Input account yourself from the dashboard or iPhone app, choosing who takes over your work in each workspace first.
Sep 25, 2026. Topics: Security, Collaboration. https://www.generalinput.com/changelog/2026-09-25-account-deletion
You can now delete your own General Input account, from the dashboard or the [iPhone app](/mobile), without contacting support. Before anything is removed, you see every workspace you belong to and everything you own in each one, and you decide what happens to it.
### How it works
- **A full inventory.** Settings lists each workspace and what you own there: specialists, workflows, apps, skills, connections, databases, file stores, and browsers.
- **Choose a successor, or delete.** For each shared workspace, pick who owns your work after you, or delete it all. If you're the workspace's owner, you also pick its next owner.
- **Solo workspaces go with you.** A workspace where you're the only member is deleted with your account and erased within 30 days.
- **Preview first.** You see everything affected before confirming, and nothing is removed until you do.
- **Leave without deleting.** Leaving a single workspace offers the same choices for that workspace while keeping your account.
To remove an entire workspace instead, see [workspace deletion](/changelog/2026-08-05-workspace-deletion).
## Business plans now start at one seat
> The five-seat minimum on Business Pro and Max is gone. Buy seats for exactly the members you have, from one to 50, with prorated additions.
Sep 25, 2026. Topics: Pricing. https://www.generalinput.com/changelog/2026-09-25-business-plans-one-seat
The Business plan used to require at least five seats, so a team of two paid for five. That minimum is gone. Choose a seat count starting from one, and each seat adds its credits to the workspace's monthly allowance plus a spot for one member. See plans on the [pricing page](/pricing).
- **Pro:** $30 per seat per month, with 4,800 credits per seat.
- **Max:** $100 per seat per month, with 16,000 credits per seat.
Both tiers cover up to 50 members, and seats added later are prorated.
### How it works
- **One upgrade screen.** Choosing a plan moved out of onboarding into a dedicated upgrade screen with a seat picker that prices each plan at the count you choose. The public pricing page renders the same plan cards.
- **Onboarding starts with credits.** The plan question became a welcome-credits step: a new workspace starts with credits to build and run with, and no card is needed.
- **One page when the balance runs out.** A workspace with nothing left to spend sees a single screen offering pay-as-you-go credits, a plan, or a call with us, then returns to where it was headed.
- **Billing on the web.** The iPhone app shows your plan and balance without asking you to pay; low-balance reminders go by email and Slack.
## geni CLI and MCP server now manage specialists and runs
> geni CLI 0.19.0 and the MCP server gain specialist, file, template, sharing, credential, version, and run-control operations for coding agents.
Sep 25, 2026. Topics: Developers. https://www.generalinput.com/changelog/2026-09-25-cli-mcp-specialists-sharing-runs
With specialists, the `geni` CLI and the MCP server picked up a long list of operations that used to need the dashboard. Agents inside General Input got the same operations as tools, so an outside coding agent, the CLI, and a specialist all work through one set of actions. The developer docs were audited against the deployed tools and every `/v1` route in the same release. See the [developer page](/developers).
### What's new
- **Specialists.** List, create, update, delete, and pin specialists, and move resources between them.
- **Specialist files.** List, pull, push, and remove files. Every write must present the version it read, so a stale write is refused instead of overwriting someone else's change.
- **Export, import, and templates.** Export a specialist as a ZIP with chosen files and memory, import one, and install or publish specialist templates.
- **Sharing.** Inspect, grant, and revoke access for one member or the whole workspace.
- **Credential maintenance.** Check health (status, expiry, and scopes, never secrets), refresh, rename, annotate, and see what depends on a credential.
- **Run control.** Resume a paused run, follow up on a finished one, cancel, or retry with the original input, and inspect or replay trigger activity.
- **Versions.** List and revert resource versions, enable or disable resources, and capture app screenshots at phone and desktop widths.
### Breaking changes
MCP tools named for projects were renamed for specialists, for example `list_projects` to `list_specialists`.
## Home inbox notifications and iPhone push alerts for runs
> Notices now live in the Home inbox beside conversations and paused runs, the iPhone app gets push notifications, and failed runs alert you once.
Sep 25, 2026. Topics: Workflows, Mobile. https://www.generalinput.com/changelog/2026-09-25-inbox-and-push-notifications
Notices used to sit behind a bell, apart from the conversations and runs they were about. They're now rows in the Home inbox: unread ones appear in Unread on the dashboard and in Activity on the iPhone app, next to conversations and paused runs, and every notice is kept in History. Alongside the move, the [iPhone app](/mobile) gained push notifications, and workflows learned to tell you when a live run fails or stops to ask you something.
### How it works
- **Per-member channel switches.** Settings has one switch per event and channel (email, Slack, push), grouped into Specialists, Operations, Billing, Digests, and Sharing, the same on the dashboard and the phone.
- **Push that respects what you've seen.** A push waits five seconds and is sent only if the item is still unread, so a conversation you're watching never buzzes your phone.
- **Failed runs, announced once.** When a live workflow run fails, its owner gets an inbox notice, email, and push linked to the run. Notices stop at five per workflow per day, and failures caused by our own deploys don't notify.
- **Paused runs reach every channel.** A paused run notifies on every connected channel by default; email and Slack wait a minute, so a question you answer right away never reaches your inbox.
### Also shipped
- Emails about a run name its specialist as the sender, like "General Input - Invoice Chaser."
## Saved cloud browsers with a power button and lasting logins
> Your saved browser is now a cloud computer you switch on. Sign-ins are refreshed every three days, and agents try single sign-on before asking you.
Sep 25, 2026. Topics: Agents. https://www.generalinput.com/changelog/2026-09-25-saved-browsers-power-button
Signing an agent into a website used to run through a guided flow where an agent checked and steered each sign-in. It was slow and brittle, so we replaced it with something plainer. On the Browser page (in beta), your saved browser is a computer with a power button: switch it on, browse and sign in to anything the way you would on your own machine, and switch it off. Agents, workflows, the CLI, and the MCP server reuse whatever you signed in to.
### How it works
- **Saves as you go.** The browser saves its state on each page load and every five seconds, so a dashboard refresh reconnects to the same session. It powers off after two minutes with nobody watching, or 15 minutes of inactivity.
- **Real logins only.** A site appears in your saved logins only when a sign-in cookie or token actually appears, not just because you visited.
- **Sign-ins kept fresh.** Every three days a background visit reopens each active login's site so it reissues its cookies before common one- and two-week timeouts.
- **Single sign-on first.** A saved login remembers its provider, like "signed in with github.com." At a sign-in wall, an agent clicks that provider's button once; anything beyond that waits for you.
- **Handing a sign-in back.** When an agent hits a wall it can't pass, it sends you a card that opens your browser on that exact page.
## Capacity: per-workspace rate limits and smoother schedules
> An audit of what would break at ten times our usage lifted three limits: per-workspace rate limits, non-blocking token refreshes, and spread-out schedules.
Sep 24, 2026. Topics: Platform. https://www.generalinput.com/changelog/2026-09-24-capacity-improvements
We audited what would break if usage grew tenfold, sized each fix from measured numbers, and lifted three limits that were already slowing busy workspaces down. The result is steadier performance at the moments workspaces are busiest, like the top of the hour when many schedules fire at once.
### How it works
- **Per-workspace rate limits.** Calls from workflows to our built-in services shared one rate limit across every workspace. Each workspace now gets its own 1,000 requests a minute, so one busy workspace can't slow down another.
- **Token refreshes no longer block the database.** Refreshing a connected account's token used to hold a database connection during the provider's response. A token that's still valid now skips the lock entirely, waiting requests release their connection while they wait, and provider calls time out after 30 seconds.
- **Schedules spread out.** Each scheduled trigger starts at a small, stable offset, so a burst of runs on the hour no longer spills past the pool of warm sandboxes.
- **Faster long conversations.** Looking up the latest message in a long conversation dropped from 67 ms to 0.4 ms.
### Also shipped
- Saving an unchanged schedule no longer rewrites it, which used to drop its next pending run.
- 31 read-only triggers now retry temporary server errors.
## Real Postgres, MySQL, and SQL Server errors for AI agents
> Agents now see the database's own error message instead of a generic failure, so they can fix their own SQL queries on Postgres, MySQL, and SQL Server.
Sep 24, 2026. Topics: Integrations, Agents. https://www.generalinput.com/changelog/2026-09-24-database-error-messages
When a query against your [PostgreSQL](/apps/postgresql), [MySQL](/apps/mysql), or SQL Server database failed, agents used to get a generic "Database query failed" with no detail. That hid real SQL mistakes, like a misspelled column or a missing join, from the agents that could have fixed them. Now the agent sees the database's own message and corrects the query itself.
### How it works
- **Classified at one boundary.** A rejected query returns the database's own message, a statement timeout and an unreachable server each get their own status, and a refused login is reported as such.
- **Clear connection errors.** When you connect a database that refuses or can't be reached, you see the server's own reason instead of a generic failure.
- **Platform faults read as platform faults.** Internal errors and overloads in our hosted databases are reported as service problems, not as mistakes in the agent's query.
Your hosted [workspace databases](/changelog/2026-03-26-workspace-databases) get the same treatment.
## Hosted Token Tunnel: your AI plan, no desktop app needed
> Sign in to your Claude, ChatGPT, Gemini, or Grok subscription once in Settings and use it in every workspace, with no computer left running.
Sep 22, 2026. Topics: Models, Security. https://www.generalinput.com/changelog/2026-09-22-hosted-token-tunnel
Running agents on your own AI subscription used to mean keeping General Input Desktop open on a computer that stayed online. The hosted [Token Tunnel](/token-tunnel) runs that connection in our cloud instead. In Settings, under AI connections, sign in to Claude, ChatGPT, Gemini, or Grok once, and those models appear in the model picker in every workspace where you're an active member. New members can connect a subscription during onboarding, and the iPhone app has the same page.
### How it works
- **Sign in on the provider's own pages.** Sign-in runs in a hosted browser framed by the dashboard, using the provider's standard secure flow, and only the provider's own pages can open.
- **One isolated connection per person.** Each person's connection runs in its own isolated environment, walled off from everyone else's, with credentials encrypted and bound to their owner.
- **Several subscriptions stack.** Requests rotate across every account you connect for a provider, and a rate-limited account fails over to another within the same request. When all are limited, the error says when they reset, and Auto routes to another model.
- **Accounts leave only on the provider's word.** An account is removed only when the provider rejects it, and you get an email.
- **Updates don't cut streams.** In-flight chats finish before a host is replaced, and every proxy update must answer a real chat on all four providers before it ships.
## Claude Opus 5.5, GPT-6 Sol and Luna, and Grok 4.7 arrive
> Opus 5.5 and GPT-6 replace Opus 5 and GPT-5.6 at lower per-token prices, Grok 4.7 replaces 4.6, and Auto tiers span four providers.
Sep 22, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-09-22-opus-5-5-gpt-6-grok-4-7
We moved the catalog to the newest generation of each frontier model family, and saved selections followed on their own: a workflow, app, or specialist that names a retired model now runs on its successor, so nothing had to be re-saved. Subscription models through Token Tunnel follow the same upgrades. Compare everything on the [models page](/models).
### What changed
Prices are per million input and output tokens.
- **Claude Opus 5.5** replaces Opus 5 and 4.8 at $4 / $20, down from $5 / $25, with a 1M-token context.
- **GPT-6 Sol** replaces GPT-5.6 Sol at $2 / $10, down from $5 / $30.
- **GPT-6 Luna** replaces GPT-5.6 Luna at $0.10 / $0.50, down from $0.20 / $1.20. Both GPT-6 models have a 1.05M-token context.
- **Grok 4.7** replaces Grok 4.6 and 4.5 at the same $2 / $6, with a 500K context and a new xhigh effort level.
- **Claude Sonnet 5.5** replaced Sonnet 5 on September 28.
- **Default effort per model.** Every model used to default to medium; each now has its own default within its supported range.
### Auto
Each tier starts cheap on the cost-to-intelligence frontier and falls back across four providers (OpenAI, Anthropic, Google, xAI), so one provider's outage can't take a tier down. When one of your own subscription models is online, Auto uses it in place of the tier's cloud model.
## Token Tunnel detects a sleeping computer in seconds
> Runs on a subscription model served from General Input Desktop now move on in about a second when the computer sleeps, instead of up to 4 minutes.
Sep 16, 2026. Topics: Models, Platform. https://www.generalinput.com/changelog/2026-09-16-token-tunnel-fault-detection
When a run used a model served from your own computer through [Token Tunnel](/token-tunnel) and that computer slept or lost its network, the run could wait 90 seconds to 4 minutes before failing. Fallback then tried every other model on the same unreachable machine before reaching a cloud model. In a typical week, about 96% of device disconnects were only caught by a server ping, mostly computers that fell asleep with the tunnel still marked connected.
### How it works
- **Faster detection.** The desktop app closes its connection before the computer sleeps, which fails open requests in about a second, and a dropped connection is detected within 15 seconds instead of 30 to 45.
- **Disconnects aren't model faults.** Timers are tuned per provider and can never fire before dead-connection detection, so an unreachable computer is always reported as a disconnect.
- **No retrying a dead machine.** When the device is unreachable, fallback skips every other model it serves and goes straight to the next option.
- **Native endpoints.** Claude runs over the Messages API and ChatGPT's Codex models over the Responses API, with tool inputs streamed as they're written. A 32 KB file write used to mean 99 seconds of silence; the longest gap is now 1.6 seconds.
- **Thinking isn't a stall.** The desktop proxy sends a keepalive while a model thinks quietly, so long reasoning is no longer cut off.
### Also shipped
- Workflows pinned to a subscription model are no longer disabled when one account hits a rate limit.
## GPT-6 Astra and Claude Fable 5.1 available on hosted keys
> GPT-6 Astra and Claude Fable 5.1 join the hosted catalog for every workspace, GLM 5.3 and DeepSeek V4.1 Flash arrive, and long-context pricing ships.
Sep 15, 2026. Topics: Models, Pricing. https://www.generalinput.com/changelog/2026-09-15-gpt-6-astra-claude-fable-hosted
GPT-6 Astra and Claude Fable 5.1 are now hosted models: any workspace can run them on General Input's provider keys without connecting a subscription. Until now both were reachable only through [Token Tunnel](/token-tunnel). Compare them on the [models page](/models).
### What changed
- **GPT-6 Astra:** a 1,050,000-token context, at $10 per million input tokens and $50 output, with cache reads at $1. Above 272,000 input tokens, input bills at 2x and output at 1.5x.
- **Claude Fable 5.1:** a 1,000,000-token context and up to 128,000 output tokens, at $10 input and $50 output, with cache reads at $0.25.
- **Same-model failover.** Both are also served through Vercel AI Gateway and OpenRouter, so an overload moves the same model to another route before any fallback.
- **Long-context pricing** is now a catalog field: an input threshold plus multipliers. Each request in a long run is priced against its own context size.
### Also shipped
- GLM 5.3 replaces GLM 5.2, and GLM 5.3 Flash joins with a 1M-token context.
- DeepSeek V4.1 Flash replaces V4 Flash, adds vision, and costs half as much: $0.22 input and $0.66 output per million tokens.
- Token Tunnel now delivers instructions to Claude and ChatGPT subscriptions as a real system prompt.
## Transfer ownership and offboard members without orphans
> Owners can hand any workflow, app, skill, project, or credential to another member, and deactivating someone transfers or deletes what they own.
Sep 15, 2026. Topics: Collaboration. https://www.generalinput.com/changelog/2026-09-15-ownership-transfer
Every resource in a workspace can now change hands, and deactivating a member can no longer strand what they built. An owner can give a workflow, app, skill, project, credential, database, store, or browser to any active member. When someone leaves, an admin moves everything they own to a successor, or deletes it, in the same step as the deactivation.
### How it works
- **Three kinds of move.** A transfer gives one resource away, a handoff moves everything a member owns to a successor, and an adopt lets an admin take over resources whose owner is already inactive.
- **Workflows move with their wiring decided.** For each connection the giver owns, they choose share or disconnect. Transfer stays disabled until every choice is made, because a workflow wired to something its new owner can't see breaks on its first run.
- **Projects move with their contents**, with a preview of what will move. Other members' resources in the project stay theirs.
- **No lost races.** Every owner change checks the expected current owner first.
- **Deactivation takes a policy.** Transfer everything, delete everything, or neither only when the member owns nothing, all in one transaction.
### Sharing surfaces
- Every resource names its owner, and sharing reads as Private or Shared.
- Controls you can't use are disabled with a tooltip that says why, never hidden.
### Also shipped
- Enable or disable a workflow or skill from the agent, the API, or `geni resource enable`.
## Exactly-once triggered runs, and cancels that keep output
> Run starts go through a transactional outbox, cron and webhook firings are deduplicated, and cancelling a run keeps the step the agent was writing.
Sep 14, 2026. Topics: Workflows, Platform. https://www.generalinput.com/changelog/2026-09-14-exactly-once-triggers
Three gaps at the edges of a run are closed. A run that exists in the database is always handed to a worker, a redelivered schedule or webhook can no longer execute twice, and stopping a run no longer throws away the text the agent was in the middle of writing.
### How it works
- **Durable starts.** Starting, resuming, or continuing a run writes its queue job in the same database transaction as the run itself, through an outbox. A publisher sends committed jobs immediately and a one-second sweep repairs anything it missed. If publishing keeps failing, the run fails with that reason instead of sitting queued forever.
- **One run per firing.** Every triggered run carries a key unique to its trigger firing, so a redelivered event finds the existing run instead of starting a second one.
- **Cancel keeps output.** A worker saves its partial step before reporting cancelled, and a run that finished in the meantime isn't overwritten.
### Also shipped
- Rerun and Test are no longer disabled while another run of the same workflow is in progress.
- Apps queue requests during busy moments (up to 256 in flight and 2,048 waiting) instead of rejecting anything past 16.
## Full-screen file viewer for PDFs, images, and spreadsheets
> Files in chats and workflow runs open in a full-screen viewer: zoomable images, continuous PDFs, spreadsheets with sheet tabs, and playable media.
Sep 14, 2026. Topics: Agents, Workflows. https://www.generalinput.com/changelog/2026-09-14-file-viewer
Agents produce and read a lot of files: reports, spreadsheets, screenshots, recordings. Opening one now fills the screen with the file itself, on a black backdrop with just a close button and a download button, on the dashboard and in the iPhone app alike.
### How it works
- **Images** fit the screen; double-click to zoom and drag to pan.
- **PDFs** scroll continuously as white pages and render only the pages near what you're reading, so a long document stays fast.
- **Spreadsheets** keep their grid, with tabs for each sheet.
- **Word documents** sit on a white page, and **HTML files** render as themselves, so an email draft looks the way it will in an inbox.
- **Markdown, code, slides, and archives** each get a clean page, and **audio and video** play straight from storage.
- **Run inputs and outputs too.** The files a workflow run received and produced open in the same viewer from the run's side panel, with the preview chosen by file type.
## Request protection, run-scoped credentials, and new roles
> Requests from untrusted origins are rejected, runs reach only the credentials they declare, and model and billing management get their own permissions.
Sep 14, 2026. Topics: Security, Collaboration. https://www.generalinput.com/changelog/2026-09-14-request-and-credential-hardening
A security review of the request surface, and of how runs reach credentials, shipped as a set of protections. None of them change how a workflow is written; they narrow what a request or a run can reach.
### How it works
- **One trusted-origin list.** Cross-origin settings, sign-in, file storage, the OAuth popup, session handling, and live connections all read the same list. A state-changing request from an untrusted origin is rejected.
- **Run credentials bound to the declaration.** A run can use only the credentials, databases, and stores its own slots declare, and the database proxy checks that a credential belongs to the run before connecting.
- **Validated connection fields.** An OAuth connection accepts only the fields an integration declares, such as an instance URL, each checked against its allowed pattern before anything is stored.
- **Retry-safe top-ups.** A credit purchase carries a request ID, so a retried submit can't charge twice.
- **Clearer connect links.** The hosted page that completes a connection names the workspace and the person who asked.
### New permissions
Manage workspace models, and Manage subscription and seats. Both default to Admins and Owners and are enforced on the server as well as in settings.
## New dashboard Home with pins, one feed, and a Resources page
> The dashboard opens on Home: pinned work in the sidebar, one feed of chats and runs waiting on you, and one Resources grid instead of four lists.
Sep 10, 2026. Topics: Specialists, Workflows. https://www.generalinput.com/changelog/2026-09-10-home-and-resources
The dashboard is rebuilt around what you come back to. The new Home became the dashboard for every workspace in early September: one feed of your work, your pins in the sidebar, and every workflow, app, and skill in a single Resources grid, so you no longer need to remember which list page something is on. (Projects here are what we now call specialists.)
### How it works
- **Pins.** The sidebar holds pinned workflows, skills, apps, projects, and chats in an order you drag. A pinned workflow opens to its launch form and runs, a pinned skill to a chat with that skill attached, and a pinned app to the app itself.
- **One feed.** Home lists chats, editing sessions, and runs waiting on your answer, newest first, built once on the server. The unread count and the inbox are the same list filtered, so the number and the list can't disagree.
- **One Resources page.** Every workflow, app, and skill in one grid, with kind, ownership, and sort filters plus search. Old links redirect.
- **Resource art.** Listings draw each resource's own art: an app's thumbnail, or a workflow's trigger and the integrations it calls.
### Also shipped
- One actions menu per resource kind (enable, rename, copy, export, move, pin, delete), shared everywhere the resource appears.
## Version-checked publishing stops overwritten workflows
> Every publish from an agent, the geni CLI, or MCP is checked against the version it started from, so a stale draft is refused instead of reverting work.
Sep 10, 2026. Topics: Workflows, Developers, Collaboration. https://www.generalinput.com/changelog/2026-09-10-publish-conflict-checks
Two chats editing the same workflow could silently revert each other: a conversation holding a draft from days earlier could publish it and wipe out changes another chat had made since. Every publish from an agent, the `geni` CLI, or an MCP client is now conditional on the version its working copy was taken from. A stale copy is refused with instructions to re-read and merge, never applied.
### How it works
- **Every working copy knows its base.** A chat's draft or a CLI directory records the version it came from, and one rule decides the check for every publish path.
- **No bypass.** Holding no working copy doesn't skip the check, and the only override is an explicit force by a person.
- **The winner's files are safe.** The version is claimed before files are written, so a losing publish can never clobber the winner.
- **Unpublished work is kept.** An untouched draft refreshes itself when a run starts; a draft with edits is kept and its publish is checked instead.
- **Refusals are recovery steps.** Read what's live, combine both sets of changes, publish again.
### How we tested it
344 tests cover the change, including an end-to-end replay of the original incident against a real database and real stored files, with the key checks mutation-tested.
Older versions of the `geni` CLI are asked to upgrade.
## Multi-provider model routing with same-model failover
> A model can now be served by several ranked provider routes. On overload, runs move the same model to its next route before switching models.
Sep 8, 2026. Topics: Models, Platform. https://www.generalinput.com/changelog/2026-09-08-model-route-failover
Model routing now lives in one engine, and a model can be served by more than one route. Every chat model in the catalog has its native provider plus routes through Vercel AI Gateway and OpenRouter. When a provider is overloaded, a run first moves the same model to its next untried route, and only then falls back to a different model, so a provider incident doesn't have to change which model is doing your work.
### How it works
- **Ranked routes.** Native lab APIs (Anthropic, Vertex, OpenAI, xAI, Fireworks) first, then Vercel AI Gateway, then OpenRouter.
- **Route hops are free.** A run tracks the routes and the models it has tried separately, so switching routes doesn't use up a model fallback.
- **Caching follows the route.** Each route declares how it handles prompt-cache breakpoints, reasoning effort, and session affinity, so a Claude request through a gateway keeps its cache discounts.
- **Embeddings too.** Embedding requests use the same routes, and a route returning vectors of the wrong size is rejected so stored vectors stay comparable.
### How we verified it
Live suites run every route of every model at every effort level, plus streaming with tools, vision, and cache reads. They caught a provider alias that had started returning errors in production, which is now pinned to its dated version.
### Also shipped
- Muse Spark 1.3, the first model served only through a gateway.
## Scoped API keys and platform keys for the General Input API
> Every /v1 route now requires one named scope. Personal keys act as you; workspace platform keys let outside agents use docs, credentials, and services.
Sep 3, 2026. Topics: Developers, Security. https://www.generalinput.com/changelog/2026-09-03-scoped-api-keys
The public `/v1` API now has two kinds of key with separate, explicit scopes, and every route checks exactly one. A personal key acts as you and manages your workflows, skills, memory, projects, and triggers. A platform API key, now available to every workspace, acts for the workspace: it can search the integration catalog, hand out credential tokens, and call hosted services, but can never start a run or touch members' resources. That split lets you give an outside agent, such as Claude Code, Codex, or your own, exactly the access it needs. See the [developer page](/developers).
### How it works
- **No wildcards.** Platform keys hold docs, vault, and platform scopes; personal keys hold workflow, skill, memory, project, and trigger scopes.
- **Hosted services.** A single endpoint reaches internet search, scraping, email, image generation, and chat completion, billed to the workspace.
- **Credentials.** List the credentials a key's member can use (field names, never values), or issue a fresh access token through the same audited path runs use. Refresh tokens are never returned.
- **Connections.** Open a pending connection that a person completes through the regular sign-in flow.
- **Tied to a person.** If the member who created a platform key leaves, the key stops working.
### Also shipped
- A full `/v1` reference in the developer docs, including scopes, errors, and rate limits.
## Workflow filters skip irrelevant events before agents run
> An optional filter.ts runs before a workflow, decides in code whether a run happens, and ends rejected events as filtered instead of waking an agent.
Sep 3, 2026. Topics: Workflows, Agents. https://www.generalinput.com/changelog/2026-09-03-workflow-filters
Any code or agent workflow can now carry an optional `filter.ts`: a cheap, deterministic check that decides whether a run happens at all, and on what input. A busy webhook or poll trigger no longer wakes an agent for every event. A code check costs roughly a tenth of a cent; an agent run costs 40 to 250 times that before it does any work, so on a noisy trigger the filter decides most of the bill.
The same mechanism is also a custom trigger: paired with a schedule, a filter can poll any source that has no built-in trigger, skip what it has already seen, and run only on new items. The trigger picker has a new Custom option that writes the schedule and filter for you.
### How it works
- **Same context as the workflow.** The filter runs in the sandbox with the workflow's credentials, settings, databases, and firewall rules, and returns skip with a reason, or run with an input.
- **Filtered is a success.** Skipped events end with a filtered status, never count toward auto-disable, and show their reason in history.
- **Fails open.** An error, a 60-second timeout, or no decision means the workflow runs anyway; a filter is an optimization, never access control.
- **Evidence in the document.** The filter block shows 7-day counts, the last five rejections, and a "too strict?" warning after 25 events with no passes.
## Integration docs tested against live APIs every day
> A test suite per service checks what our integration docs claim against the real API daily, so agents hit fewer failed calls. Plus August's new apps.
Aug 31, 2026. Topics: Integrations, Platform. https://www.generalinput.com/changelog/2026-08-31-integration-docs-verified-live
Agents call your integrations from our documentation of each API, so a wrong field name or a retired endpoint becomes a failed run. We now check that documentation against the live APIs. Each service gets a test suite whose probes encode what our docs claim, and a daily run replays them. The finding we care about is a faithful probe failing because the API disagrees with the docs.
### How it works
- **Two modes.** Verify replays the committed suite deterministically, with no model involved, so it's cheap enough to run daily. Author runs a maintainer agent to write or repair a suite, but never in response to a failure, which would hide real drift.
- **Evidence, not status codes.** A 200 can carry an error body and a 403 can mean a higher plan tier, so each result is classified as passed, failed, needs a human, or requires a paid plan.
- **Safe test data.** Probes may create the records they need, but never spend money, message a person, or touch data they didn't create.
- **Scope today.** The catalog documents 34,054 operations across 557 services, and 201 services with test accounts put 11,169 operations in reach.
### Also shipped
- New integrations in August include 1Password, Bitwarden, SignWell, DocuSeal, Hootsuite, Sprout Social, GoHighLevel, Paylocity, Snowplow, and Circleback, 16 with polling triggers. See the [catalog](/apps).
## Team roles: custom member roles that your apps enforce
> Define your own roles, like Sales Rep or Finance, assign them to members, and apps you build show or allow things based on each person's roles.
Aug 31, 2026. Topics: Collaboration, Apps. https://www.generalinput.com/changelog/2026-08-31-team-roles
Workspaces can now define their own team roles, such as Sales Rep, Recruiter, or Finance, and give each member one or more of them. Apps use those roles to decide what each person sees and can do, so one app can serve a whole team with the right view for each person.
### How it works
- **Your vocabulary.** Team roles are names your workspace defines. They sit alongside, and never replace, the built-in Owner, Admin, and Member roles that control workspace permissions.
- **Assigned centrally, applied by each app.** You manage who holds which role in one place; each app decides what a role means for it, for example "Finance can approve refunds."
- **Delivered to every app.** App handlers, app agents, and app screenshots all receive the viewer's roles, and an app can read the live member roster instead of keeping its own copy.
- **Built into authoring.** The app builder can see your members and their roles while it builds, so you can ask for role-based behavior in plain language.
- **Permissioned.** Changing someone's roles requires permission to update member roles.
This builds on [apps knowing who's using them](/changelog/2026-07-20-apps-warm-servers-viewer-identity).
## Projects: one home for a goal's workflows, apps, and skills
> Projects, now specialists, group the workflows, apps, skills, files, and chats behind one goal, with sharing that flows to everything inside.
Aug 28, 2026. Topics: Specialists, Collaboration. https://www.generalinput.com/changelog/2026-08-28-projects
Projects shipped to every workspace. Projects, which have since become specialists, give a goal that takes several pieces one home: a skill holding the procedure, a workflow running it on a schedule, an app where someone reads the output, and the reference files behind all three. Start a chat inside a project and everything it builds lands there. Read the [Projects announcement](/blog/introducing-projects).
### How it works
- **One home per resource.** A workflow, app, or skill lives in at most one project, and you can move work in or out at any time.
- **Chats follow the work.** A chat that creates a project moves into it, and @-mentioning a project attaches its inventory to your message.
- **Access flows down, and the narrower setting wins.** Share a project as viewer, runner, or editor and the role reaches everything inside, while a resource can still carry its own narrower setting.
- **Chats stay private.** A conversation in a shared project belongs to the person having it; what the project knows lives in its memory and files.
- **Project memory.** Each project keeps its own memory, refreshed daily from its files, workflows, runs, and chats.
- **Everywhere you work.** Sidebar, search, @-mentions, the agent's tools, the MCP server, and the CLI.
Access rules are tested in CI against a real Postgres database: 90 tests cover every role, the workspace boundary, and every way of revoking access.
## Self-updating AI memory: Memory refreshes itself every day
> The knowledge base is now Memory: skill bundles every agent loads, revised daily by a memory agent that reads what actually happened in your workspace.
Aug 26, 2026. Topics: Agents. https://www.generalinput.com/changelog/2026-08-26-memory-daily-sweep
The knowledge base is now Memory, one name across the product: a Memory page in the dashboard, and workspace and personal memory available to every agent. The bigger change is that it no longer depends on someone typing facts into it. A memory agent runs daily for each memory, reads what happened since its last run, and revises it: confirming what still holds, changing what changed, adding what's new, and removing what's now wrong.
### How it works
- **Signals, not prompts.** Each run reads the team's chats, what was built, and what ran. If nothing happened, the run is skipped.
- **Requests, not notes.** What you type into the memory form, and what agents file with the `remember` tool, becomes a request the memory agent consumes on its next run. People's requests are instructions; agents' inferences are checked against the evidence first. No agent edits memory directly.
- **Shared memory is never about a person.** Workspace memory records how the business works and never stores names, emails, or phone numbers; facts about one person belong in their personal memory.
- **It does its own research.** The workspace memory agent reads what the company has published, starting from its website, and dates every claim so later runs know what to re-check.
- **A new page.** A timeline of everything that shaped the memory sits beside the memory itself.
## Workflows declare databases and browsers as credential slots
> Workspace databases and saved browsers are chosen in credential slots like integrations, and a workflow run can reach only the resources it declared.
Aug 24, 2026. Topics: Workflows, Security, Integrations. https://www.generalinput.com/changelog/2026-08-24-databases-browsers-credential-slots
Workflows, apps, and templates declare the integrations they need as credential slots. Databases and browsers now go through the same slots: the picker lists your databases and browsers beside your integration credentials, and publishing refuses a slot nobody filled. The result is that a workflow states up front every resource it touches, and the platform holds it to that list.
### How it works
- **One decision point.** A single service decides whether a slot is backed by an integration credential or a platform resource, so pickers, validation, publishing, duplication, templates, and app handlers all treat both the same way.
- **Keys scoped to the declaration.** The short-lived key a workflow run receives carries the resources it declared. A call to a database outside that set is refused with a message naming the slot to add, so an agent can correct itself.
- **Browsers only where declared.** A saved workflow can open only the browsers it declared, so every browser session traces back to a browser someone can point at.
- **Built-in services read as ready.** Services we host, like internet search, show as usable without asking you to connect them.
- **Migrated first.** Existing workflows that used a database or browser were updated to declare it before strict scoping turned on.
## Auto runs on your Token Tunnel subscriptions first
> With Claude, ChatGPT, Grok, or Gemini connected through Token Tunnel, Auto now runs on those subscriptions first and falls back to hosted models last.
Aug 19, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-08-19-auto-routes-to-token-tunnel
Auto picks the model for each request. It now considers the models you've connected through [Token Tunnel](/token-tunnel) and prefers them: when a suitable one is online, Auto runs there, on the subscription you already pay for, instead of on a hosted model billed in credits. A Default to local models switch in the Auto menu controls it; it's on by default once you have a connected model.
### How it works
- **Tiers.** Auto classifies a request as an answer, an execution, or a build, mapped to Fast, Balanced, or Max. Anything it can't classify goes to Max.
- **A ladder per tier.** Each tier has an ordered list of Token Tunnel models, and every tier includes all four providers (Claude Code, Codex, Grok, Gemini), so a single subscription is enough.
- **Never a blocker.** If the switch is off, nothing in the ladder is online, or a lookup fails, the tier runs its normal hosted fallbacks.
- **Every subscription before the cloud.** When one subscription hits its limit mid-run, the run moves to the next online one, and falls back to a hosted model only after all of them are exhausted.
### Also shipped
- Smaller models that couldn't reliably finish agent work became opt-in, and anyone defaulted to one moved to Auto.
- Gemini 3.7 Flash replaced Gemini 3.6 Flash, and Settings, Models now shows who hosts each model and where.
## Weekly digest email of your workspace's AI activity
> Every member gets a weekly digest of their workspace's activity: runs, chats, runtime, and a team leaderboard, with starter ideas for quiet weeks.
Aug 19, 2026. Topics: Collaboration. https://www.generalinput.com/changelog/2026-08-19-weekly-digest-email
Since August 19, every member gets a weekly digest of what happened in their workspace: how much work their workflows and agents did, and how they compare with the rest of the team. It arrives on Wednesdays, covers the previous Monday through Sunday, and comes once for each workspace you belong to.
### How it works
- **Your numbers.** Workflow runs, chat messages, AI runtime, and tokens for the week, calculated the same way as the [Analytics](/changelog/2026-07-29-workspace-analytics-member-reports) page so the two never disagree.
- **A team leaderboard.** In workspaces with two or more members, the digest ranks the top five by runtime and shows where you stand if you're outside them.
- **Quiet weeks get ideas instead.** If a week was light, you get a starter digest with suggestions for what to try next rather than a page of zeros.
- **A notification, not marketing.** The digest reports your own activity, so it follows your notification settings (Settings, Notifications, Digests) and also appears in the app.
- **Sent once.** Each digest is claimed before it's sent, so nobody gets the same week twice.
## Browser use: AI agents that use websites with saved logins
> Agents now drive a real hosted browser for sites without an API, reuse per-site logins you save once, and pause so you can sign in from chat.
Aug 18, 2026. Topics: Agents, Integrations. https://www.generalinput.com/changelog/2026-08-18-browser-use-saved-logins
For work that has no API (a supplier portal, an internal tool, a retailer's product page), agents can now open a real hosted browser. Every member has a browser that keeps per-site logins: sign in to a site once, and later runs open already signed in. When a run lands on a sign-in page, it pauses, you sign in through a live view in the chat, and the run picks up with the result. Agents never ask for a password and never type one.
### How it works
- **Scripts, not clicks.** The agent writes and runs a Playwright script against the hosted browser, so one step can complete a whole browsing task instead of a model round trip per click.
- **Logins you can trust.** Each browser lists its sites and whether each is signed in. Before a login is saved, it's proven by loading a page that requires an account.
- **Watch and take over.** The live browser appears in the chat while the agent works, and you can take control.
- **Unattended runs fail clearly.** Saved workflows can't request logins; an expired one reports that it needs reconnecting.
- **Ground rules.** Page text is data, never instructions. Agents buy, send, post, or delete only when you asked for exactly that, and treat a CAPTCHA as a stop.
- **Metered by the minute** on top of sandbox time.
### Also shipped
- `geni browser` brings the same hosted browser to scripts on your machine.
## General Input MCP server for Claude and ChatGPT
> Connect Claude, ChatGPT, Claude Code, or any MCP client to General Input and get the full agent toolset: every integration, a sandbox, and building.
Aug 18, 2026. Topics: Developers, Integrations, Agents. https://www.generalinput.com/changelog/2026-08-18-mcp-server-full-agent-toolset
Connect once from Settings, then MCP, and Claude, ChatGPT, Claude Code, Codex CLI, or any other MCP client can work in your General Input workspace. The connector isn't a read-only window or a short list of popular actions: it mounts the full toolset our own agent uses, so the model in your AI tool can call any operation across your connected integrations, build and edit workflows, apps, and skills, run them, and read their traces. Read the [MCP announcement](/blog/general-input-mcp).
### How it works
- **Per-person sign-in.** The server URL names your workspace, and everyone signs in with their own account through OAuth, so each person gets exactly their own access.
- **One toolset, two surfaces.** The connector registers the same tools the in-product agent runs, so the two can't drift.
- **A persistent sandbox.** Each member's connector session gets a warm Linux sandbox that keeps its files between calls and is released after five minutes idle. Read-only calls never start one.
- **Credentials stay in the vault.** Secrets are injected inside the sandbox, scrubbed from output, and every use is audit-logged with the reason the model gave.
- **Interactive cards** where the client supports MCP Apps, like a Connect button for a missing integration.
- **Billing.** Reasoning runs on your AI tool's own model; sandbox time and platform services bill to the workspace.
## Per-seat Business plan pricing: $30 Pro or $100 Max per seat
> Self-serve plans are now one per-seat Business plan: Pro at $30 with 4,800 credits per seat, or Max at $100 with 16,000. Added seats are prorated.
Aug 17, 2026. Topics: Pricing. https://www.generalinput.com/changelog/2026-08-17-per-seat-business-plan
We replaced the flat self-serve plans with one per-seat Business plan in two tiers. Pro is $30 per seat per month and funds 4,800 credits per seat. Max is $100 per seat per month and funds 16,000 credits per seat. Both started at 5 seats and go up to 50; larger teams go through Enterprise. The free tier is unchanged. See the [pricing page](/pricing) for current plans.
### How it works
- **Seats are bought, not counted.** At checkout you pick a tier and a seat count. That number sets what you pay, how many members the workspace can hold, and how many credits arrive each month.
- **Membership changes never touch the invoice.** Inviting or removing someone doesn't change what you pay. Add seats on the plan card and the prorated difference is charged immediately, with the matching credits granted right away.
- **Existing subscribers keep their terms.** Workspaces on a retired plan renew against the plan they bought.
- **A calculator on the pricing page.** Choose a tier and drag the seat count to see monthly credits and estimated capacity in runs per day and hours saved, using a conservative 500 credits per hour.
## Embedded agents: launch AI agents from buttons in your apps
> Apps can declare agents anyone using the app launches from its interface, each a full agent workflow with its own sandbox, history, and approvals.
Aug 12, 2026. Topics: Apps, Agents. https://www.generalinput.com/changelog/2026-08-12-agents-embedded-in-apps
Apps can now declare background agents that someone using the app launches from a button: a Research button on each row of a lead table, a Draft reply button on each support ticket. An app agent is not a reduced version of a workflow agent. It follows the same agent workflow spec, runs in its own sandbox, and gets the same run history, streaming, approval steps, and billing. Ask for one in the chat panel next to your app, the same way you'd ask for a new chart. Read the [launch post](/blog/agents-inside-your-apps).
### How it works
- **Same files as an agent workflow.** Each agent lives in the app bundle with its inputs (the button's variables), model and time budget, and a written procedure.
- **The interface is the trigger.** App agents have no triggers of their own, and their connections must be the app's own, so the Integrations tab shows which agent uses each.
- **Owner's access, viewer's identity.** Anyone the app is shared with can launch an agent; it runs through the owner's connections and knows who launched it.
- **Guardrails.** Every launch is confirmed, since runs spend credits, and an app can run at most 25 agents at once. Agents write results back to the app's data.
- **Tested by the builder.** The app-building agent runs its own agents and reads their traces before handing the app over.
### Also shipped
- App containers ship with more libraries preinstalled, from data tables to Excel, Word, and PDF generation.
## Send messages to an agent mid-run without stopping it
> Message an agent while it works and your note lands at its next step, so it adjusts in the same run. Undo takes it back before it lands.
Aug 11, 2026. Topics: Specialists, Agents. https://www.generalinput.com/changelog/2026-08-11-queued-messages-mid-run
You no longer have to wait for an agent to finish before you correct it. Send a message while a run is in flight and it shows up as a dimmed Queued bubble. At the agent's next step, the message is added to the conversation and the agent responds to it in the same run. Changed your mind? Undo on the bubble, or Escape in the composer, takes it back and returns the text and attachments to the composer. Queuing works in chat and while building workflows, apps, and skills.
### How it works
- **Held apart until it lands.** A queued message is stored separately, so nothing that reads the conversation (the model, summaries, titles, exports) sees it early. It joins the conversation exactly once.
- **One entry point.** Every way a run starts or restarts checks the queue, and mid-run the agent checks it at each step boundary.
- **Nothing stranded.** If a run fails or is cancelled, queued messages become ordinary history; if it pauses for your input, the queue waits for the resumed run.
- **Undo is unambiguous.** Either the model never saw the message, or it's already in the transcript.
## Getting Started checklist that ticks itself off as you go
> A Getting Started checklist in the sidebar walks new members through six first steps and marks each one done the moment you actually do it.
Aug 8, 2026. Topics: Specialists. https://www.generalinput.com/changelog/2026-08-08-getting-started-checklist
New members now get a Getting Started checklist in the dashboard sidebar. It walks through six first steps: connect an integration, chat with the agent, install [Token Tunnel](/token-tunnel), build a workflow, build an app, and build a skill. Each step opens a short explanation with a button that takes you straight to the right page.
### How it works
- **It notices what you did.** Steps complete themselves when you actually do them, like connecting your first account or publishing your first app, so there's nothing to tick by hand. You can still mark a step done or finish them all.
- **Progress never goes backward.** Once a step is complete it stays complete, even if you later delete what you built.
- **Live updates.** The checklist refreshes every minute and whenever you return to the tab, and disappears once every step is done.
- **Clean new workspaces.** New workspaces start empty instead of seeded with example workflows, apps, and skills, so the checklist guides you to build your own.
Prefer a guided tour? The [crash course](/blog/getting-started-crash-course) covers signing up, connecting your tools, and shipping a workflow and an app.
## Workspace deletion: permanently delete a workspace
> Owners can now delete a workspace from Settings. It goes dark immediately, and its subscription, data, files, and schedules are removed for good.
Aug 5, 2026. Topics: Collaboration, Security. https://www.generalinput.com/changelog/2026-08-05-workspace-deletion
Owners can now delete a workspace from Settings. Confirm by typing the workspace's name, and it becomes invisible and inert immediately: members lose access, and nothing in it runs again. Its data is then removed permanently, not just hidden.
### How it works
- **Owner only.** Deleting a workspace requires the Owner role.
- **Outside services first.** The workspace's Stripe subscription is cancelled immediately, its hosted databases are deleted, and its stored files are swept from storage.
- **One transaction for the data.** Every record the workspace owns, from workflows and runs to connections, chats, and memberships, is deleted in a single ordered transaction, so a failure can't leave half a workspace behind.
- **Schedules stop.** Scheduled and polling triggers are removed once the deletion commits.
- **Nothing forgotten.** A test reads the database schema and fails if a new kind of workspace data is ever added without being covered by deletion.
- **Your account stays.** Deleting a workspace doesn't delete your account or your other workspaces; for that, see [account deletion](/changelog/2026-09-25-account-deletion).
### Also shipped
- New accounts verify their email address before entering the product, and accepting an invitation has a clearer screen.
## AI usage analytics per member, with PDF and CSV reports
> A new Analytics page shows who in a workspace uses AI and for what: per-member builds, runs, tokens, and credits, exportable to PDF or CSV.
Jul 29, 2026. Topics: Collaboration. https://www.generalinput.com/changelog/2026-07-29-workspace-analytics-member-reports
The new Analytics page shows who in a workspace is using AI and for what. It replaces the separate executions, usage, and security audit pages with one view, and adds a Members section that ranks every active member by what they built and ran.
### How it works
- **One page, one toolbar.** Executions (success rates, runs by workflow), usage (spend over time with receipts), and security (audit logs) share one time window, from 24 hours to 90 days, and a member filter.
- **The member table.** One row per member: the code workflows, agent workflows, apps, and skills they own, plus executions, chat messages, building iterations, tokens, and credits spent in the window. Columns sort, and pending invites appear on the workspace view.
- **Reports.** Export the table as a branded PDF or CSV. Each member also has a profile PDF covering their totals, what they've built, and a token breakdown by prompt, completion, cache read, and cache write.
- **Access.** Everyone can see and download their own row; seeing anyone else requires the usage audit permission.
### Also shipped
- Resource pages open full screen, and list pages became document-style lists with thumbnails.
- Update an API key in place instead of deleting and re-adding it, so bound workflows keep working.
- Sharing an app or skill sends its own notification with a direct link.
## Claude Opus 5 replaces Opus 4.8, and Kimi K3 arrives
> Claude Opus 5 takes over from Opus 4.8 at the same price with saved selections upgraded automatically, and open-weight Kimi K3 is now available.
Jul 28, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-07-28-claude-opus-5-and-kimi-k3
Two hosted models landed at the end of July. Compare them on the [models page](/models).
### Claude Opus 5
Opus 5 replaces Opus 4.8 in the hosted catalog at the same rates: $5 input, $25 output, $0.50 cache read, and $6.25 cache write per million tokens. Every stored reference to Opus 4.8, like a member's default model or a workflow's configured model, now resolves to Opus 5, so nothing needs reconfiguring.
- A 1,000,000-token context window.
- Reasoning effort from low to max, defaulting to medium.
- If Anthropic is overloaded, runs fall back to GPT-5.6 Sol, then Gemini 3.5 Flash.
- Opus 4.8 remains available through Token Tunnel for members whose Claude plan still serves it.
### Kimi K3
Kimi K3 is Moonshot's open-weight model, available for chat, building, workflow runs, and as a utility model.
- A 1,048,576-token context window.
- Reasoning effort from low to max, defaulting to high.
- Image input.
Both models bill on actual cache reads and writes, like the rest of the catalog. Read why we think [running open-source models yourself](/blog/chinese-open-source-models-safest-cheapest) is a safe, cheap option.
## Faster apps: warm servers, viewer identity, and deep links
> App actions run on a warm per-app server, handlers know who is using the app, and every page inside an app gets its own shareable URL.
Jul 20, 2026. Topics: Apps. https://www.generalinput.com/changelog/2026-07-20-apps-warm-servers-viewer-identity
July reworked how apps run and how people move through them. Actions respond faster because they hit a server that's already up, app code can behave differently for different members, and any page inside an app can be linked to directly. See how apps fit into [agents inside your apps](/blog/agents-inside-your-apps).
### How it works
- **Warm per-app servers.** Each app gets one warm sandbox server shared by everyone using it, instead of a fresh sandbox per action. It starts on the first action, stops after five minutes idle, picks up new versions in place, and is replaced on any failure.
- **Viewer identity.** Handlers still run with the app owner's connections, but every call now knows who is using the app, so agents can build rules like "only I can delete" or "reps see their own accounts." Checks are enforced in server code, never by hiding a button.
- **Browser bar and deep links.** Apps get back, forward, and home controls, and the current page is reflected in the dashboard URL, so a copied link opens on the same page with the same filters.
- **Stays signed in.** Long-open apps refresh their session in place, without reloading or losing state.
- **Connections first.** An app with unconnected required integrations shows a connect screen instead of loading into errors.
### Also shipped
- Handlers can stop long requests early when the person closes the page.
## Token Tunnel: use your Claude, ChatGPT, Gemini, or Grok plan
> Token Tunnel connects the AI subscriptions you already pay for, so chats and workflows run on them with no API keys and no per-token charges.
Jul 19, 2026. Topics: Models, Platform. https://www.generalinput.com/changelog/2026-07-19-token-tunnel
[Token Tunnel](/token-tunnel) is available to every workspace. Install General Input Desktop, sign in to the AI subscriptions you already have (Claude, ChatGPT, Gemini, or Grok), and those models appear in General Input as your own. Chats and agent workflows run on them without an API key, usage isn't billed per token, and your provider logins live only on your computer. Learn more about [running local and personal models](/blog/what-is-a-local-ai-model).
### How it works
- **A menu bar app.** One universal macOS build for Apple Silicon and Intel. Sign in to General Input with a device code, then to each provider with its own login. The app auto-connects and updates itself.
- **The tunnel.** The app keeps an outbound connection to our gateway. When a run uses one of your models, the request is relayed to your computer, sent to the provider on your subscription, and streamed back.
- **Model sync.** Any model in our catalog that your subscription serves appears automatically, including models we don't host, like Claude Fable 5.
- **Explicit offline handling.** If your computer is off, chat says so immediately and never silently switches to a paid model. Agent workflows choose a policy: skip the run after retrying for 15 minutes, or fall back to a hosted model you pick.
- **Clear errors.** Rate limits, expired logins, and provider outages are labeled as such, and each device serves 10 simultaneous requests by default.
## Chat streaming that survives sleep and dropped Wi-Fi
> Chats now reconnect cleanly after a laptop sleeps or the network drops, showing every message exactly once while your run keeps working.
Jul 18, 2026. Topics: Platform. https://www.generalinput.com/changelog/2026-07-18-chat-streaming-reconnects
Long agent runs used to look broken at exactly the wrong moments: a laptop waking from sleep, a train going through a tunnel, a flaky office network. The chat stream is rebuilt so reconnecting is clean. Your run keeps working on our servers the whole time, and the chat catches up exactly where it left off.
### How it works
- **Every event is numbered.** After a reconnect, the chat applies only newer events, each exactly once, so text never appears twice.
- **Fewer, larger updates.** Text is batched into about 25 updates per second instead of one per token, in strict order with tool and status updates.
- **Dead connections are caught fast.** The server checks in every 15 seconds, and 40 seconds of silence forces a reconnect even when the browser thinks the connection is fine. A network change or a tab waking up reconnects immediately.
- **An outage isn't a failure.** During a long drop, the chat keeps retrying and says the connection is out while your run continues.
- **Specific errors.** A run that ends on a provider problem now says why, such as rate limited or provider unavailable, with a next step.
### Also shipped
- One memory-hungry command no longer takes down an agent's whole sandbox; only that command is stopped, with a note to process the data in smaller pieces.
## Auto model routing: the right model tier for every task
> Auto sends each run to a Fast, Balanced, or Max model tier based on the work, falls back across providers, and is now the default.
Jul 17, 2026. Topics: Models, Agents. https://www.generalinput.com/changelog/2026-07-17-auto-model-router
Most requests don't need the most expensive model, and some need nothing less. Auto now makes that call per run. It sits at the top of the model picker with three tiers (Fast, Balanced, and Max), and new members start on Auto everywhere. Members who already chose a default keep it.
### How it works
- **Fixed surfaces route directly.** Building or editing a workflow, app, or skill always runs on Max; agent workflow executions default to Balanced.
- **Chat is classified.** Your workspace's utility model sorts each message into build, execute, or answer, mapped to Max, Balanced, and Fast. A question about your own data counts as execute.
- **Fails up, never stops.** If classification fails, the run goes to Max. Routing can never block a run.
- **Sticky while warm.** A conversation stays on its tier for five minutes so follow-ups keep hitting the same prompt cache, and the choice is stored on the run so resumes stay consistent.
- **Outage-tolerant.** Each tier has a fallback list spanning three providers.
The classifier scored 50 out of 50 across three runs on an eval built from the real (anonymized) distribution of chat openers.
### Also shipped
- Grok 4.5, and the GPT-5.6 family (Sol, Terra, Luna) with 1M-token context windows.
## 500+ integrations: Workday, NetSuite, Okta, AWS, and more
> The General Input integration catalog grew from 275 to more than 500 in July, adding enterprise ERP and HR, AWS, payments, commerce, and AI services.
Jul 17, 2026. Topics: Integrations. https://www.generalinput.com/changelog/2026-07-17-integration-catalog-passes-500
Our integration count went from 275+ at the start of July to 500+ on July 17. Each integration ships with operation-level API documentation agents search, a setup guide, a credential check on connect, and poll triggers wherever the service has a real new-items stream. Browse the full [integrations catalog](/apps).
### What was added
- **Enterprise systems:** [Workday](/apps/workday), [NetSuite](/apps/netsuite), Microsoft Dynamics 365, SAP SuccessFactors, Coupa, Odoo, Brex, Mercury, Canvas, and Clio.
- **Developer and cloud:** AWS (S3, SES, SQS, Lambda, CloudWatch, DynamoDB), Azure DevOps, CircleCI, Snyk, LaunchDarkly, [Okta](/apps/okta), Auth0, Heroku, Render, Neon, and Firebase.
- **Payments and commerce:** PayPal, Adyen, Braintree, Paddle, Chargebee, BigCommerce, eBay, and Shippo.
- **Data and AI:** Databricks, ClickHouse, Fivetran, dbt Cloud, Pinecone, Vapi, Replicate, and DeepL.
### How it works
- **Validated on connect.** Adding a credential runs a minimum-privilege check and records account context for agents; keys with restricted scopes still pass.
- **Per-tenant OAuth.** Workday, NetSuite, Dynamics 365, and others collect your tenant or region before sign-in, so calls reach the right instance.
- **Triggers start from now.** New poll triggers baseline when enabled, so turning one on never replays a backlog.
- **Guardrails.** User-supplied hosts must match the provider's domains, and keys are scrubbed from error logs.
## Export and import workflows, apps, and skills as .zip files
> Move any workflow, app, or skill between workspaces as a .zip, and upgrade template-installed workflows in place when the template improves.
Jul 15, 2026. Topics: Workflows, Developers. https://www.generalinput.com/changelog/2026-07-15-export-import-and-template-upgrades
You can now move a workflow, app, or skill between workspaces as a file. Export downloads the resource's live bundle as a .zip; import validates it, shows a preview, and creates the resource. Credentials never travel in the file, so an imported workflow arrives ready for you to connect your own accounts.
### How it works
- **Export.** Anyone with read access can export. For workflows, you choose which setting values ride along, so workspace-specific values can stay behind.
- **Import.** Upload a .zip of up to 50 MB from the Workflows, Apps, or Skills page. Unsafe archives are rejected, the bundle is validated against that resource's spec, and nothing is created until you confirm.
- **Safe arrival.** Imported workflows arrive disabled, so nothing fires webhooks or polls before you review it. Missing connections show as needing attention, and a failed import leaves nothing half-created.
- **Portable skills.** Skills use the standard Agent Skills folder layout, so they move between hosts.
### Template upgrades
Workflows installed from a template now show their source and an upgrade banner when the author publishes a newer version. Upgrading snapshots your current code, applies the new version, keeps your connections, settings, triggers, sharing, and run history, and lets you roll back.
### Also shipped
- Template authors choose which setting values ship as defaults.
## Agent runs resume after restarts without repeating actions
> Every agent tool call with side effects is journaled, so a run interrupted by a deploy resumes without sending that email or creating that record twice.
Jul 15, 2026. Topics: Agents, Platform. https://www.generalinput.com/changelog/2026-07-15-tool-call-journal
An agent run can take many minutes, and occasionally one of our servers restarts in the middle of it, for example during a deploy. Until July 15, the resumed run could repeat a step it had already taken. Now every tool call with side effects is recorded in a journal before it runs and again when it finishes, so a resumed run picks up exactly where it left off.
### How it works
- **Journal before and after.** A tool call that changes something, like sending an email or updating a CRM record, is written to the journal when it starts and when it completes.
- **Rebuilt, not repeated.** After a restart, the run rebuilds its unsaved step from the journal instead of executing those calls again.
- **Uncertain calls are verified.** A call that was in flight when the server stopped is marked as uncertain, and the agent is told to check whether it happened before trying again.
- **Reads simply re-run.** Read-only calls have no side effects, so they're safely repeated.
### Also shipped
- Long conversations with many images stay under the context limit: image reads are stored as pointers to the file instead of inline data, and counted at their real cost.
## Voice dictation in every prompt box, with smart cleanup
> Every prompt box has a mic. Speech streams to live transcription, then a cleanup pass removes filler and applies your spoken edits.
Jul 8, 2026. Topics: Specialists. https://www.generalinput.com/changelog/2026-07-08-voice-dictation
Describing a workflow out loud is faster than typing it, but raw transcripts are full of false starts. Every prompt box in the dashboard now has a mic button: words appear as you speak, and when you stop, a cleanup pass turns the ramble into the prompt you meant and puts it in the box. It never submits for you. Read [why talking to agents is faster than typing](/blog/voice-dictation-faster-than-typing-ai-agents).
### How it works
- **Streaming transcription.** Your browser streams audio directly to AssemblyAI with a short-lived token; audio never passes through our servers, and a live strip shows the partial transcript.
- **Cleanup that edits.** On stop, your workspace's utility model removes filler, keeps only the corrected version when you change your mind mid-sentence ("make it a Google Sheet, no wait, a Slack message"), and applies spoken instructions like "delete the part about revenue charts" to what's already in the box.
- **Never loses words.** If cleanup fails, the raw transcript is added instead.
- **Everywhere you type.** Chat, workflow, app, and skill building, onboarding, and the new-resource dialogs.
- **Six languages.** English, Spanish, German, French, Portuguese, and Italian, detected automatically.
Transcription costs about $0.19 per audio hour, prorated per second.
## Agent skills: reusable SKILL.md instructions for every agent
> Skills are portable SKILL.md bundles in the open Agent Skills format that any agent loads when a task matches. Build them in chat or import a .zip.
Jul 7, 2026. Topics: Agents, Developers. https://www.generalinput.com/changelog/2026-07-07-skills-workspace-resource
Workflows and apps do work; skills teach agents how to do it. A skill is a `SKILL.md` instruction bundle in the open Agent Skills format: your email conventions, a report layout, a triage procedure, plus any scripts and templates that go with it. Skills now sit beside workflows and apps with their own page, every agent in the workspace can load them, and because the format is portable, a skill written here works in other Agent Skills hosts and the reverse. See why we favor [small agents over mega-skills](/blog/small-agents-not-mega-skills).
### How it works
- **The bundle.** `SKILL.md` holds a name and description, with optional scripts, references, and asset folders for runnable code, long reference material, and templates.
- **Progressive disclosure.** Agents see only each skill's name and description until one matches the task; then the body loads and its files are placed in the sandbox. You can also @-mention a skill to use it directly.
- **Validated on publish.** Every link inside the bundle must resolve, and a failed publish changes nothing.
- **Lifecycle.** Describe a skill and an agent drafts it, or import a .zip. Every publish saves a version, skills can be switched on or off, and they share like workflows.
### Also shipped
- The `geni` CLI authors skills with the same commands as workflows and apps.
## Prompt-cache-aware pricing for every AI model
> Every model now carries cache-read and cache-write rates, so each call is priced on what the provider actually served, and balances update live.
Jul 1, 2026. Topics: Pricing, Models. https://www.generalinput.com/changelog/2026-07-01-cache-aware-model-pricing
Agent runs resend the same long prefix on every step: instructions, tool definitions, and the conversation so far. Most of it comes back from the provider's prompt cache at a fraction of the input price. Every model in the catalog now carries four rates (input, output, cache read, cache write), and each call is priced on what the provider actually served.
### How it works
- **Three input buckets.** Uncached tokens bill at the input rate, cache reads at the cache-read rate, and cache writes at the cache-write rate. On Claude Sonnet 5, for example, input is $3.00 per million tokens and a cache read is $0.30.
- **More cache hits.** Each conversation's requests carry a session-affinity key, so providers serving open-weight models route a conversation back to the replica holding its warm cache. For Anthropic models we mark the end of the history as a cache breakpoint, so earlier turns read from cache on every step.
- **Pruning only when needed.** Old tool results are trimmed only when a conversation nears its context limit, since pruning earlier would invalidate the cache.
- **Auditable.** Every usage record stores its cache-read and cache-write counts.
### Also shipped
- Credits deduct while a run is in progress, and open dashboards update the balance live.
- Workflows that keep failing because credits ran out now pause with a notice saying how to resume.
## Claude Sonnet 5 becomes the default model; Opus 4.8 arrives
> Managed Claude models now run on Anthropic's API, upgraded to Claude Sonnet 5 and Opus 4.8. MiniMax M3 and GLM 5.2 also joined in June.
Jun 30, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-06-30-claude-sonnet-5-default
On June 30 we moved our managed Claude models to Anthropic's own API and upgraded them in the same step. Claude Sonnet 4.6 became Claude Sonnet 5, Claude Opus 4.6 became Claude Opus 4.8, and Haiku 4.5 stayed as it was. Sonnet 5 is now the default model for chat and for building workflows and apps, and the agents in our built-in templates use it too. See all models on the [models page](/models).
### How it works
- **Nothing to edit.** Old model IDs resolve to the new ones, so workflows, workspace defaults, member defaults, and conversation model pickers that named Sonnet 4.6 or Opus 4.6 now run on Sonnet 5 and Opus 4.8.
- **Native effort.** Effort settings map to Anthropic's native effort parameter.
- **Pricing.** Listed at $3 input and $15 output per million tokens for Sonnet 5, and $5 and $25 for Opus 4.8. Both accept images and files.
### Earlier in June
- **MiniMax M3** replaced MiniMax M2.7, bringing a 512K-token context window and vision. Saved selections moved over automatically.
- **GLM 5.2** joined the catalog for agent work, with a 1M-token context window.
## Agent Firewall: per-operation API permissions for AI agents
> Agent Firewall lets you block reads, writes, or individual API operations per connection or workspace-wide, enforced before requests leave the sandbox.
Jun 26, 2026. Topics: Security, Integrations. https://www.generalinput.com/changelog/2026-06-26-agent-firewall
OAuth scopes and API keys are coarse. Connecting Google Sheets grants edit and delete on every spreadsheet, and an agent holding that credential inherits all of it. [Agent Firewall](/agent-firewall) moves the unit of control down to the operation: on any connection you can block whole categories (every write, every read) or individual operations, and admins can set standing rules across the workspace. Agents keep the access their job needs and lose the actions you never want them to take. Read [why scopes are the wrong unit of control](/blog/oauth-scopes-wrong-unit-of-control-for-ai-agents).
### How it works
- **Built on our integration docs.** Every documented operation carries an HTTP method, path, and read or write category, and every integration lists its domains, so rules expand into precise request matchers.
- **Enforced in the sandbox.** Outbound traffic passes through a proxy that checks each request against the rules. A blocked call gets a clearly marked 403 telling the agent which operation was denied and to ask a workspace owner.
- **Layered rules.** Workspace rules combine with each connection's own rules, so a member can't loosen a block an admin set.
- **Everywhere agents run.** The same policy applies in chat, agent workflow runs, and `geni exec` on your machine.
Find it in Settings under Agent Firewall, on each connection's page, and in the Integrations Tools tab.
## Workspace notifications delivered as Slack DMs
> Admins connect Slack once, members are matched by email, and each notification type can go to email, a Slack DM, or both.
Jun 25, 2026. Topics: Collaboration, Integrations. https://www.generalinput.com/changelog/2026-06-25-slack-notifications
Notifications used to go out by email only. A workspace admin can now connect [Slack](/apps/slack), and each member chooses, per notification type, whether it arrives by email, as a Slack DM from the workspace's bot, or both. That covers auto-disabled workflows, failed triggers, low balance, balance refills, revoked credentials, and workflows, credentials, or databases shared with you.
### How it works
- **One-click setup.** Add to Slack runs our managed Slack OAuth, or an admin can pick an existing Slack connection. The token is stored encrypted on a dedicated notification connection, separate from anyone's personal credential.
- **Automatic matching.** Members are matched to Slack users by email, admins can correct any match, and resyncs never overwrite a manual match. Members can send themselves a test DM.
- **Permissioned.** Managing the Slack connection requires a new Manage notification channels permission, held by Admins and Owners.
- **Independent channels.** Delivery fans out per channel, so a Slack failure never blocks the email, and email and Slack render from the same content.
- **Graceful failure.** An invalid token disables the workspace connection with a reconnect path, an unreachable member disables only their link, and rate limits never disable anything.
## Atomic app builds on a dedicated esbuild service
> App builds now run as queued jobs on their own service. A failed build keeps the last good version live, with build status shown as it happens.
Jun 24, 2026. Topics: Apps, Platform. https://www.generalinput.com/changelog/2026-06-24-app-builds-dedicated-esbuild-builder
Apps used to build on our API servers. An agent edit could trigger duplicate builds, concurrent builds were silently dropped, publishing cleared the running app until the next good build, and a crashed build could leave an app stuck on "building." We rebuilt the pipeline in two steps.
### Queued, atomic builds (June 3)
- Builds are queue jobs, one active build per app. Triggers that arrive mid-build collapse into a single rerun with the latest files instead of being dropped.
- Build output is written to a fresh location and the app switches to it only when the build succeeds, so a failed build leaves the previous version serving.
- The app pane follows build status (queued, building, ready, error) and swaps in the new version as soon as it's ready. Failed builds keep the compiler output for you and the agent.
### A dedicated builder (June 24)
- Builds run on their own autoscaling service, so bursts never compete with API traffic. The builder holds no credentials or database access and moves files through short-lived signed links.
- esbuild replaced Vite for bundling, and type checking reuses a warm TypeScript program, so library types are parsed once rather than on every build.
- Icon imports are rewritten per icon, cutting the icon library from about 4,500 modules to a few dozen.
## Workflow file uploads and optional settings and connections
> Workflows accept uploaded files as input, settings with defaults can be left blank, and connections can be optional so one workflow fits many tool stacks.
Jun 22, 2026. Topics: Workflows. https://www.generalinput.com/changelog/2026-06-22-workflow-file-inputs-optional-settings
Over June we loosened what a workflow needs before it can run. It can take files as input, its settings can be optional, and so can its connections, so one workflow can serve people with different tool stacks and a missing nice-to-have no longer blocks a run.
### File inputs
- An input marked as a file renders as an upload field on the run form, and the file lands in the workflow's workspace for code or agent steps to read. You can restrict the accepted type, such as PDF only.
- Multi-file fields accept several uploads at once.
- When the building agent tests a file input, it stages a real sample file, so tests match what real uploads do.
- Each run records the input it received alongside the trigger that fired it.
### Optional settings and connections
- A setting with a default can be left blank, and the run uses the default.
- A connection can be marked optional: the workflow runs either way, skipping the dependent step when the account isn't connected. Optional connections are labeled in the workflow view and in templates.
- The building agent reserves optional connections for secondary paths, never for something the core flow depends on.
### Also shipped
- Workflows with several triggers can tell which one fired.
## Build workflows and apps from Claude Code with the geni CLI
> Coding agents like Claude Code can now create, validate, publish, and test General Input workflows and apps from local files with the geni CLI.
Jun 20, 2026. Topics: Developers. https://www.generalinput.com/changelog/2026-06-20-geni-cli-workflow-app-authoring
The `geni` CLI can now author code workflows, agent workflows, and apps end to end. Local files are the draft: your coding agent edits them with its own tools, `geni resource validate` checks them without saving, `geni resource publish` makes them live, and `geni workflow test` runs the published version in the cloud. It's the same build loop the in-product agent follows, backed by the same server-side checks. See the [developer page](/developers).
### How it works
- **Lifecycle commands.** `geni resource` creates, pulls, validates, configures, and publishes any resource; `geni workflow` tests runs and inspects executions; `geni app` checks builds, runs handlers, and reads errors.
- **One validation path.** The CLI runs the same spec and type checks as the dashboard, so the two can't drift.
- **Access-checked.** Every command that takes a resource ID is checked against your membership, exactly as in the dashboard.
- **Byte-exact files.** Images and document templates survive pull and publish unchanged.
- **Full run traces.** `geni workflow execution --download` writes the run, logs, input, and agent thread to disk, the same trace the cloud agent reads when it debugs.
Developer docs cover every command.
## One AI agent on every surface, with skills loaded on demand
> The separate chat, workflow, app, and database agents are now one agent that loads authoring skills when a task or page needs them.
Jun 19, 2026. Topics: Agents, Specialists. https://www.generalinput.com/changelog/2026-06-19-one-agent-progressive-skills
We ran four agents: one for chat, one that built workflows, one that built apps, and one for database questions. Each had its own prompt, tools, and run lifecycle, so a fix to one often missed the others. They're now a single agent, and what changes between surfaces is the subject it's pinned to (home, a workflow, or an app), which decides its instructions, preloaded skills, and tools.
### How it works
- **Skills on demand.** Every conversation starts with the same base tools. Authoring tools sit behind skills, loaded when needed or automatically when you open a workflow or app, and they persist for the conversation.
- **Shared tools.** Per-type tools merged into resource-wide ones that behave the same everywhere.
- **Test runs in chat.** Tests stream live inside the conversation as an expandable card.
### Quality from evals
- New code and app eval suites, built from the same prompts the New Workflow and New App dialogs send, drove a round of prompt and scaffold fixes.
- Code workflows are type-checked on the server against the real runtime library before they're saved.
- A run stops with a clear message when the same tool hits the same error five times in a row.
### Also shipped
- Web search falls back across three providers when one is rate-limited.
- Interactive runs are capped per member and queue visibly instead of competing for workers.
## Global search across workflows, apps, connections, and files
> Search everything in a workspace from one box: workflows, apps, connected accounts, databases, members, and context files, with fuzzy matching.
Jun 11, 2026. Topics: Platform. https://www.generalinput.com/changelog/2026-06-11-global-search
As workspaces filled up, finding something meant remembering which page it lived on. Global search, added June 11, puts one search box in the dashboard header that looks across the whole workspace at once.
### How it works
- **Everything in one place.** Results cover workflows and apps, connected accounts, databases, members, and context files, each marked with its own icon so you can tell them apart at a glance.
- **Forgiving matching.** Search matches names and descriptions with fuzzy, typo-tolerant matching, and boosts results that start with what you typed, so the most likely match rises to the top.
- **Connections by what they're for.** Connected accounts match on their title, description, notes, and service, so searching "warehouse" finds the database connection you annotated that way.
- **People too.** Members match by name or by email address.
Combined with [notes on connections](/changelog/2026-03-19-connect-postgres-mysql-sql-server), it's the fastest way to find the right account for a task, for you and for the agent.
## AI agent memory: a summary of your workspace and you
> A memory agent keeps a short summary of your company and of you that every agent run loads, and opens detailed files only when needed.
Jun 7, 2026. Topics: Agents, Specialists. https://www.generalinput.com/changelog/2026-06-07-workspace-and-personal-memory
We rebuilt workspace and personal context into a memory system. Each workspace and each member now has a distilled summary of at most 200 words covering identity and working style: who you are, how you write, and the standing instructions you've given. Every agent run loads that summary, so agents know the essentials on every run without reading your whole knowledge base each time.
### How it works
- **Summary plus files.** Facts, processes, and reference material stay in context files, indexed by title and description, and agents open them only when a task calls for it.
- **One memory agent.** It seeds a nearly empty knowledge base with public research about your company, fills gaps in a sparse one, and distills a large, messy one into a few focused topic files. When notes conflict, the most recent thing you said wins.
- **Trustworthy sources.** Web content is treated as data, never as instructions. Researched claims carry their source, and personal facts come only from you and your connected data.
- **Refreshes on its own.** Your edits mark the summary stale, and a background refresh runs once changes have been quiet for a few minutes. You can also refresh on demand or turn automatic refresh off.
### Also shipped
- A redesigned Context page with Workspace and Personal views, an autosaving editor, and live refresh status.
## Custom integrations: connect any API-key or public REST API
> When an app isn't in the catalog, an agent turns its API docs into a private integration that works in chat, workflows, and apps.
Jun 3, 2026. Topics: Integrations, Agents. https://www.generalinput.com/changelog/2026-06-03-custom-integrations
Until now, an API outside our catalog meant filing a request and waiting. You can now add it yourself: in Integrations, choose New, then Custom Integration, and describe the API or drop in its docs. An agent drafts a private integration for your workspace, and once published it works everywhere a built-in integration does. Agents call it in chat, and workflows and apps declare it as a connection.
### How it works
- **Same structure as ours.** A custom integration has a description, an authentication definition, common docs (base URL, auth header, pagination, rate limits, errors), a setup guide, and one documented operation per endpoint, so agents call it as reliably as a catalog integration.
- **Reads any format.** Drop in an OpenAPI or Swagger spec, a Postman collection, a PDF, or plain docs, or just give a name or link and the agent researches the public reference.
- **API keys or public APIs.** Define one or more key fields and mark which are secret; secret fields are encrypted, scrubbed from output, and injected only at run time.
- **Publish when complete.** Publishing requires authentication, docs, a setup guide, and at least one operation; the agent tells you what's missing.
- **Workspace-owned.** Members connect their own credentials by following its setup guide.
Services that only support OAuth can still be requested through Integration Request.
## Apps: build internal tools and dashboards by describing them
> Describe a dashboard, lookup, or one-click action and an agent builds a working React app on your connected data, opened right from your workspace.
Jun 2, 2026. Topics: Apps. https://www.generalinput.com/changelog/2026-06-02-apps-internal-tools
Workflows run in the background, but some work needs a screen. Apps, available to every workspace since June 2, are single-purpose internal tools an agent builds for you: a dashboard over your CRM, a quick lookup across two systems, a one-click action your team runs every day. Describe what you need, and the app builder writes it, tests it against your real accounts, and publishes it to your workspace.
### How it works
- **A real app.** Each app is a React interface rendered in a sandboxed frame, with server-side handlers that call your connected integrations using your stored credentials. Handlers run isolated in a sandbox and are metered by runtime.
- **A component library.** Apps start from a shared set of tables with sorting and filter chips, tabs, detail views, and settings screens, plus CSV and PDF export.
- **Built and checked by the agent.** The builder runs handlers against the live app, reads its render errors and handler logs, and confirms each build before handing it over. Handler results can be cached for 15 minutes to keep apps fast.
- **Not a workflow.** Apps have no triggers or schedules; they connect only the integrations they need.
Apps have grown a lot since, from [warm servers and per-viewer identity](/changelog/2026-07-20-apps-warm-servers-viewer-identity) to [agents inside your apps](/blog/agents-inside-your-apps).
## Hour-long workflow runs and automatic sandbox failover
> Every run can now last 60 minutes, deploys finish in-flight runs instead of cutting them off, and runs fail over between sandbox providers.
May 30, 2026. Topics: Platform, Agents. https://www.generalinput.com/changelog/2026-05-30-hour-long-runs-and-sandbox-failover
Long runs used to stop for reasons that had nothing to do with the work. Code workflows were capped at 5 minutes and chat and scheduled runs at 10, a deploy could stop a worker mid-run, and a sandbox provider having a bad hour meant failed runs. Over May we lined all of those up.
### How it works
- **One 60-minute limit.** Chat, scheduled runs, resumes, follow-ups, and code workflows share a single 60-minute timeout. A run that hits it is recorded as timed out, separately from a cancel, and the error states the limit.
- **Deploys drain instead of killing.** On shutdown, a worker stops taking new jobs and finishes the ones it has, for up to an hour, so a deploy no longer cuts off a long run or its final file save.
- **Autoscaling.** The worker pool scales with load, waiting before scaling up so brief spikes don't overreact and before scaling down so bursts can finish.
- **Two sandbox providers.** Sandboxes run on E2B or a second provider, built from one shared image so every run sees the same packages, fonts, and tools.
- **Failover.** If the primary provider errors or isn't ready within its deadline (10 seconds by default), the run moves to the next provider automatically.
## Per-model reasoning effort, GPT-5.5, and Gemini 3.5 Flash
> The model picker shows an estimated price and an effort control using each provider's native levels. GPT-5.5, Gemini 3.5 Flash, and Kimi K2.6 arrived.
May 28, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-05-28-model-effort-and-new-models
Choosing a model is two decisions: which model, and how hard it should think. The model picker now shows both, along with what each option costs, and five new models arrived during May. Compare them on the [models page](/models).
### How it works
- **Estimated price.** Models are grouped into Basic, Capable, and Advanced tiers, each showing an estimated price in credits per 1,000 tokens, weighted 10 to 1 toward input because agent turns read far more than they write.
- **Native effort levels.** The effort picker lists only the levels the model's provider accepts (low to xhigh for GPT, low to max for Claude Sonnet and Opus, minimal to high for Gemini Flash) and passes the value straight through.
- **Saved defaults.** Your default model is saved with its effort, separately for chat and for building workflows.
- **Switch mid-conversation.** Conversations can move between providers, with tool-call history normalized so the next provider accepts it.
### New models
GPT-5.5, Gemini 3.5 Flash, Kimi K2.6, Qwen 3.6 Plus, and GLM 5.1. Most open-weight models moved to Fireworks AI, and saved selections moved with them, so no workflow had to be reconfigured.
## Self-serve plans: sign up and start building, no waitlist
> Anyone can now sign up and start free with 800 credits, or subscribe to a paid plan through Stripe for discounted monthly credits.
May 26, 2026. Topics: Pricing. https://www.generalinput.com/changelog/2026-05-26-self-serve-plans
Until now, a new workspace waited on a waitlist or a sales conversation before it could run anything. Sign-up is now self-serve: pick a plan during onboarding and start building. Everyone still on the waitlist was approved when self-serve went live. Current plans are on the [pricing page](/pricing).
### Plans at launch
- **Free:** 800 credits to start, pay-as-you-go top-ups, one member, no card required.
- **Individual and Team:** a monthly credit grant at a discount to pay-as-you-go, for solo builders and for teams that share workflows and credentials.
- **Enterprise:** custom pricing, unlimited members, and on-premise deployment.
### How it works
- **Credits live in grants.** Every credit comes from a subscription payment, a top-up, or a promotion. A plan's discount is applied once, when money comes in, so a credit is worth the same wherever it came from.
- **Spend order.** Usage draws from the grant that expires soonest, so monthly subscription credits are spent before purchased top-ups. Subscription credits don't roll over; top-ups never expire.
- **Seats.** Seat limits count pending invites as well as active members.
- **Stripe.** Checkout and subscription management run through Stripe.
## General Input is CASA Tier 2 verified
> General Input passed CASA Tier 2, the App Defense Alliance security assessment Google requires for restricted Gmail and Drive scopes.
May 24, 2026. Topics: Security, Platform. https://www.generalinput.com/changelog/2026-05-24-casa-verified
General Input is now CASA Tier 2 verified. CASA (Cloud Application Security Assessment) is the App Defense Alliance's review, and Google requires it before an app can use restricted scopes such as reading and modifying Gmail or full Google Drive access. To prepare, we ran the platform through the same dynamic and static scans the assessment uses and fixed what they found across the API, the dashboard, and our edge.
### What changed
- **Error responses.** Server errors return a generic message; details are logged, never echoed back.
- **Logs.** Tokens, cookies, and secrets are redacted before a log line leaves the process.
- **Caching.** Every API response is marked no-store, since responses are private to a session.
- **Rendered output.** HTML produced by agents renders in a sandboxed frame with no access to the dashboard.
- **Transport and headers.** TLS 1.2 or later, plus a Content Security Policy, HSTS with preload, clickjacking protection, a strict referrer policy, and a Permissions-Policy that restricts 21 browser features.
- **Dependencies.** Packages with published advisories are pinned to patched versions.
### Also shipped
- Password reset by email.
## Collaborative workflow editing with live presence
> Shared workflows gain an Editor role, live avatars for who's viewing, and version checks that stop one save from overwriting another.
May 21, 2026. Topics: Collaboration, Workflows. https://www.generalinput.com/changelog/2026-05-21-workflow-editor-role-and-presence
Two weeks after sharing launched, a workflow can be shared with Editor access. Editors can change instructions, triggers, credentials, settings, and agent configuration, and can ask the agent to make changes. Once more than one person can edit the same workflow, collisions become possible, so this release also makes concurrent editing safe.
### How it works
- **Version checks.** Every workflow carries a version number, and each change is sent with the version the browser last loaded. If someone else saved first, you see "This workflow was changed" with a reload option, and your unsaved edits stay on screen so you can copy them.
- **Live updates.** Each save is broadcast to everyone viewing the workflow, and open tabs refresh only when they're actually behind.
- **Presence.** The workflow page shows an avatar for each member who has it open.
- **One agent at a time.** When another member starts the agent, everyone viewing attaches to that run and watches it stream, instead of starting a second one.
- **Admin access.** Workspace admins now get Editor access to every workflow by default.
## Persistent agent workspaces that survive restarts
> Every conversation and workflow run now has a persistent work folder, so agents resume with their files instead of starting over.
May 19, 2026. Topics: Agents, Workflows. https://www.generalinput.com/changelog/2026-05-19-persistent-agent-workspaces
An agent's files used to last only as long as one sandbox. Close the conversation, or follow up on a finished run, and its scratch files, downloaded data, and half-written scripts were gone. Now every conversation and every execution has a persistent work folder, and agents run code the way an engineer does: with bash.
### How it works
- **Saved and restored.** The folder is saved to storage about once a minute after changes settle, and restored when the next turn or run starts, so a crash loses at most about a minute of work. Workspaces hold up to 1 GiB, and binary files like spreadsheets and PDFs round-trip byte for byte.
- **Follow-ups keep context.** A follow-up on a finished agent run, or a run resuming after a question, continues with the previous files.
- **Bash as the primitive.** Agent workflow scripts run directly in shell, Python, or JavaScript, receive credentials as environment variables, and call platform services over HTTP.
- **Faster first answers.** The sandbox boots in the background while the model starts responding; only the first tool call waits.
### Also shipped
- Uploaded files stay on disk with a short reference in the conversation, so a large CSV no longer overflows the model's context.
- ffmpeg, ImageMagick, and YouTube tools come preinstalled in the sandbox.
## Over 100 new integrations and 344 workflow triggers
> May added more than 100 integrations and grew the trigger library from about 60 triggers on 20 services to 344 on 52.
May 17, 2026. Topics: Integrations, Workflows. https://www.generalinput.com/changelog/2026-05-17-integrations-and-triggers-expansion
More than 100 services gained connection support in May, taking the catalog past 240 integrations: data tools ([Snowflake](/apps/snowflake), Segment, Mixpanel, Amplitude), work apps (Pipedrive, ClickUp, Zoom, Xero), web data (Firecrawl, Tavily, Bright Data), voice (ElevenLabs, Deepgram), enrichment, and media monitoring. Over the same month, the trigger library grew from about 60 triggers on 20 services to 344 on 52, so more workflows can start from the event itself instead of a schedule that checks for it. Browse them in the [integrations catalog](/apps).
### New trigger services
- [GitHub](/apps/github) (14 triggers), [HubSpot](/apps/hubspot), [Salesforce](/apps/salesforce), [Zendesk](/apps/zendesk), Freshdesk, Front, [Linear](/apps/linear), Sentry, PostHog, [Calendly](/apps/calendly), Outlook, OneDrive, SharePoint, monday.com, [QuickBooks](/apps/quickbooks), Todoist, and YouTube.
- Recruiting and HR (Greenhouse, Workable, BambooHR, Rippling) and meeting recorders (Fathom, Grain, tl;dv, Read AI).
- Custom cron expressions alongside the minute, hourly, daily, weekly, and monthly presets.
### How it works
- **Resume from a watermark.** Triggers built on change feeds save their position after each successful check and continue from it, so no event is skipped or repeated.
- **Introspection.** Every new integration reads back the account and resources it can reach, so agents build against what exists.
- **One connection layer.** Every connection now runs through our own OAuth and credential system.
## geni CLI: your connected apps for Claude Code and Codex
> Install the geni CLI and Claude Code, Codex, or Gemini CLI can call any service you've connected, with credentials injected and scrubbed from output.
May 11, 2026. Topics: Developers, Integrations. https://www.generalinput.com/changelog/2026-05-11-geni-cli
`geni` is a command-line tool that gives an AI coding agent on your machine access to the integrations you've connected in General Input. Ask Claude Code to post in a Slack channel or pull last week's Stripe charges, and it finds the right credential, reads the operation's docs, and makes the call, without ever seeing the token. Install it from npm as `@general-input/cli` and see the [developer page](/developers).
```sh
curl -fsSL https://generalinput.com/install.sh | sh
geni login
geni skills install
```
### How it works
- **Browser sign-in.** `geni login` runs a device-code flow: approve in the browser, pick a workspace, and the CLI saves a session token.
- **Discovery first.** Commands list your credentials and each operation's exact request shape and environment variable names, so the agent doesn't guess URLs.
- **Secrets injected, then scrubbed.** `geni exec` resolves each credential in the cloud, injects it as an environment variable for one local command, and redacts it from the output.
- **Every use is logged.** Each credential must be paired with a reason, which lands in your workspace's credential access log.
- **Skills for your agent.** `geni skills install` teaches Claude Code, Codex CLI, Gemini CLI, and VS Code Copilot Chat how to use it.
## Share workflows and credentials with view and run roles
> Share workflows and connected accounts with members or the whole workspace. Runners use your credentials without ever seeing the secret.
May 7, 2026. Topics: Collaboration, Workflows. https://www.generalinput.com/changelog/2026-05-07-share-workflows-and-credentials
Until now, a workflow and every credential behind it belonged to the one member who built it. You can now share both the way you share a document: pick members or give the whole workspace access, then choose what they can do. The person who connected [Salesforce](/apps/salesforce) can hand a workflow to the team, and the team runs it without their own admin login. Read the [announcement](/blog/share-workflows-and-credentials).
### How it works
- **Workflow roles.** Viewers can open a workflow and its full run history; Runners can also launch it. Editing and managing access stay with the owner.
- **Credential sharing.** A shared credential can be selected in the recipient's own workflows, but the key or token stays encrypted and is only injected into the sandbox at run time, so nobody it's shared with can read it.
- **Scoped to the workflow.** A runner doesn't need a separate credential share; the credentials a shared workflow uses resolve for that workflow's runs only.
- **Enforced on the server.** Every action that changes a workflow or its runs checks access on the server, not just in the interface.
### Also shipped
- Make a copy of any workflow you can read; credentials you don't own are removed from the copy.
- Email notifications when something is shared with you, and runs record who started them.
## Failing triggers now pause instead of disabling workflows
> Poll triggers now retry transient errors and judge failures by HTTP status. When one gives up, only that trigger pauses and the workflow stays on.
Apr 30, 2026. Topics: Workflows, Integrations, Platform. https://www.generalinput.com/changelog/2026-04-30-failing-triggers-pause-not-disable
Poll triggers check a service on a schedule and start a run for each new item. When polling broke, we used to disable the whole workflow after three consecutive failures, whatever the cause, so a provider outage could switch off an automation its other triggers were still feeding. As of April 30, a failing trigger pauses only itself.
### How it works
- **Retries in place.** Network errors and timeouts get up to three attempts within a check, and an authentication error retries with a freshly loaded credential in case the token rotated.
- **Judged by status code.** Lost access (401, 403) or a deleted resource (404, 410) pauses the trigger after one failed check. Server errors are tolerated for 25 consecutive checks, other client errors for 3. A rate limit never pauses a trigger.
- **Only the trigger pauses.** The workflow stays enabled, its other triggers keep firing, and you get a notification naming the trigger and the reason. Editing the workflow after fixing the cause turns it back on.
- **A shorter lookback.** Checks look back one hour instead of 24.
### New triggers
[Google Sheets](/apps/google-sheets) row updates, [Airtable](/apps/airtable) records, [Jira](/apps/jira) issues, [DocuSign](/apps/docusign) envelopes, Ashby candidates, Ramp transactions, [Asana](/apps/asana) tasks, and PandaDoc documents.
## Prompt library: ready-made starting points for workflows
> Browse starter prompts by use case and integration, in the product and on the web, and turn one into a code or agent workflow in a click.
Apr 30, 2026. Topics: Workflows. https://www.generalinput.com/changelog/2026-04-30-prompt-library
Knowing what to automate first is often the hardest part. The prompt library, launched April 30, collects ready-made starting points by use case and by integration. Pick one and the builder starts from that prompt, so the first draft of a workflow is a click away instead of a blank page. The same library is public at [generalinput.com/prompts](/prompts).
### How it works
- **Lighter than templates.** A prompt describes the job rather than shipping a finished workflow, so the builder adapts it to your own tools and data as it builds.
- **Your choice of workflow type.** Turn a prompt into a code workflow for fixed steps, or an agent workflow for work that needs judgment.
- **Browse by use case or app.** Filter by category, and every integration's page lists the prompts that use it.
- **Part of onboarding.** Since May, new workspaces finish onboarding with a suggested first prompt.
### Also shipped
- App prompts in August: starting points for internal tools and dashboards, not just workflows.
## Human-in-the-loop approvals for agent workflows
> Agent workflows can pause at moments you define, ask for approval or a missing detail, and resume where they left off once you answer.
Apr 28, 2026. Topics: Workflows, Agents. https://www.generalinput.com/changelog/2026-04-28-agent-workflows-ask-before-acting
Some steps shouldn't run unattended: sending a draft to a customer, choosing which leads to contact, filling in a detail only you know. Agent workflows can now stop and ask. The run parks with a Paused status, the question appears on your hub and on the execution's page, and once you answer, the agent picks up exactly where it stopped.
### How it works
- **Opt-in per workflow.** You write when to pause, for example "Pause to confirm the draft email before sending," and the agent only gets the pause tool when pausing is on.
- **One question per item.** Three posts needing approval means three questions you page through; the run resumes only when every open question has an answer.
- **No question without a decision.** If a run finds nothing to approve, it finishes without asking.
- **State survives the wait.** The agent's files are saved while paused, and waiting time doesn't count against the run's time limit.
### Where you answer
The hub replaced the home page on April 22, with a Running panel and a Waiting for you panel that collects every open question. Executions open inside the workflow page, with a link you can share.
### Also shipped
- Stop any active execution; stopped runs get their own Cancelled status.
## Onboarding interview that teaches agents your company
> New workspaces start with a short interview that writes shared and personal context, so every agent knows your company, your role, and your style.
Apr 25, 2026. Topics: Specialists. https://www.generalinput.com/changelog/2026-04-25-context-onboarding-interview
Agents now carry what they know about you from one conversation to the next. Context holds two sets of files: workspace context every member shares, and personal context only you can see. On April 25 we made filling it in the first thing a new workspace does.
### The setup interview
- **Workspace interview.** It asks what the workspace is for, looks up public information like your website, product, and brand so you don't retype it, and plays back what it found for you to correct. It writes a General profile and, when there's a visual identity, a Design file that agents use for PDFs, decks, and emails.
- **Personal interview.** After you connect your integrations, it reads a snapshot of each account (Slack channels, Linear teams, GitHub repos, Gmail labels), asks about your role and working style, and writes your profile.
- **Free during onboarding.** Later updates from the Context page bill like any other run.
### How agents use context
Every agent run gets an index of context files with a one-line description each. Agents always read the General profiles and open other files only when a task needs them. Agents write to context when you ask them to remember something, and only Admins and Owners can change workspace context.
## Usage in credits, plus an Audit section for spend and access
> Balances now show in credits at 4,000 for $25, and a new Audit section breaks down credential access, spend, and run outcomes.
Apr 23, 2026. Topics: Pricing, Security. https://www.generalinput.com/changelog/2026-04-23-credits-and-audit
Billing now shows credits instead of dollars. The rate is fixed at 160 credits per dollar, so $25 buys 4,000 credits, and refills come in packs of 1,000, 2,000, 4,000, and 10,000. Every execution and agent run shows its cost as a receipt. See current plans on the [pricing page](/pricing).
### The Audit section
Three views cover your own activity or, with permission, the whole workspace, over the last 1, 3, or 7 days:
- **Security:** every time a credential was used, the service it belongs to, what used it, and why.
- **Usage:** spend grouped by category, such as AI tokens and platform services, with receipts linking each charge to the run or chat that caused it.
- **Executions:** completed and failed runs per workflow, with a success rate.
Each workspace-wide view has its own permission, so you can grant spend visibility without granting security logs.
### Also shipped
- Background AI work, like titling chats and summarizing long conversations, appears as its own category so it's easy to tell apart from your agent runs.
## Agent skills for PDF, Word, Excel, and PowerPoint files
> Agents load skills on demand to produce finished PDFs, Word documents, Excel workbooks, slide decks, on-brand designs, and email that renders everywhere.
Apr 23, 2026. Topics: Agents. https://www.generalinput.com/changelog/2026-04-23-document-skills
Ask an agent for a deliverable and you used to get Markdown or a hand-built HTML file. Agents now have skills: packaged instructions plus the tools to produce a specific kind of output, shared across chat, the workflow builder, and agent workflow runs.
### The skills
- **PDF:** extract text and tables, merge, split, watermark, run OCR on scanned pages, and create new PDFs.
- **Word, Excel, PowerPoint:** create and read .docx, .xlsx (with formulas and styling), and .pptx files, and convert any of them to PDF.
- **Brand guidelines:** pull a company's palette, fonts, and logo from its website and apply them consistently.
- **Email design:** HTML email that holds up in Outlook, dark mode, and Gmail's 102 KB clipping limit.
### How it works
- **Loaded on demand.** Agents see a one-line description of each skill and load the full instructions only when a task needs them, which keeps prompts short and cacheable.
- **Ready to run.** The sandbox ships with the document toolchain (pypdf, python-docx, openpyxl, python-pptx, pandas, LibreOffice, Tesseract) and 48 fonts preinstalled, so the first PDF in a run doesn't wait on installs.
- **Inline previews.** Files and images an agent creates render right in the chat.
## One-click integrations with scopes you choose
> Microsoft 365, Google Workspace, HubSpot, Salesforce, and dozens more connect in one click, with read, read-and-write, or custom scopes you pick first.
Apr 14, 2026. Topics: Integrations, Security. https://www.generalinput.com/changelog/2026-04-14-one-click-oauth-and-scopes
Connecting tools was the slowest step before a first workflow could run. Dozens of providers now connect in one click: pick the service, approve the provider's own consent screen, and you're done, with no app to register and no token to paste. That covers the Microsoft suite (Outlook, [Teams](/apps/microsoft-teams), OneDrive, SharePoint, Excel), the Google suite ([Gmail](/apps/gmail), Drive, Calendar, [Sheets](/apps/google-sheets), Docs), [HubSpot](/apps/hubspot), [Salesforce](/apps/salesforce), [Airtable](/apps/airtable), Box, Dropbox, [GitHub](/apps/github), [Linear](/apps/linear), QuickBooks, Shopify, and Stripe. Read the [release post](/blog/one-click-integrations-release).
### How it works
- **Scopes you choose.** Before authorizing, pick Read, Read and Write, or Custom, with each scope described in plain language. You can edit a connection's scopes later.
- **Agents respect them.** Before calling an API, the agent checks the granted scopes and asks you to adjust the connection instead of making a call that fails.
- **Healthier connections.** Simultaneous token refreshes no longer break single-use refresh tokens, disconnecting one Google app keeps the others connected, and deleting a credential turns off the workflows that depend on it.
### Also shipped
- More than 20 new integrations, including [Jira](/apps/jira), [DocuSign](/apps/docusign), [Asana](/apps/asana), Ashby, Bullhorn, PandaDoc, GitLab, and ZoomInfo.
## Choose your workspace's AI models, or bring your own
> A Models settings page controls which AI models a workspace can use, and custom models can point at any OpenAI-compatible endpoint you host.
Apr 4, 2026. Topics: Models. https://www.generalinput.com/changelog/2026-04-04-workspace-models
Workspaces now decide which AI models their agents use. A new Models page in settings lists every managed model by provider and lets admins turn each one on or off. Teams that already run an approved model in their own cloud can add it as a custom model, which is the bring-your-own-model path for regulated or cost-sensitive teams. See the current lineup on the [models page](/models).
### How it works
- **Two roles.** Agent models appear in the model picker for chat, building, and workflow runs. One utility model handles background work like chat titles and summarizing long conversations, where a fast, cheap model is the right choice.
- **Custom models.** Any OpenAI-compatible API: a base URL, model name, context window, and capability flags. Keys are encrypted at rest.
- **Checked before anything runs.** Enabling an agent workflow confirms its model is turned on for the workspace.
- **New models.** OpenAI's GPT-5.4, mini, and nano joined Claude, Gemini, and MiniMax, and DeepSeek V4 Flash and Pro arrived on April 24 with 1M-token context windows.
### Also shipped
- Your default model is saved per member, so it follows you across browsers and devices.
## Agent workflows: AI that follows your written procedure
> A new agent workflow type hands each run to an AI agent that follows your written procedure and makes judgment calls fixed code can't.
Apr 2, 2026. Topics: Workflows, Agents. https://www.generalinput.com/changelog/2026-04-02-agent-workflows
Until April, every workflow was TypeScript that ran the same steps each time. That suits moving data between apps, and breaks down on work that needs judgment: deciding which stories in a feed matter, triaging a reply, drafting a response. Workflows now come in two types. Code workflows work as before. Agent workflows hand each run to an AI agent that follows a written procedure, and you can switch a workflow between the types.
### How it works
- **Three files.** The same manifest code workflows use (credentials, triggers, inputs), a plain-English runbook, and an agent configuration with the model, a time limit, and a reasoning effort.
- **Scripts as utilities.** The builder writes a small helper script for each external call, and the agent decides when to run them. If a script breaks mid-run, the agent can fix it and the fix is kept.
- **A living runbook.** After each test, the builder writes what worked (endpoints, response shapes, working code) back into the procedure, so later runs follow it instead of exploring again.
- **Run history as context.** Agents can pull a workflow's recent executions to debug failures against real runs.
### Also shipped
- An Activity calendar of past and upcoming scheduled runs, and per-type email alerts.
## Shopify, Cal.com, Granola, Gong, and Fireflies triggers
> Workflows can now start from Shopify orders, Cal.com bookings, Granola notes, Gong calls, Fireflies transcripts, and Zoho CRM records.
Mar 31, 2026. Topics: Workflows, Integrations. https://www.generalinput.com/changelog/2026-03-31-new-triggers-and-polling-controls
Six more services can now start a workflow on their own events, which closes loops like booking, to CRM, to meeting notes without anyone copying data by hand. A [Cal.com](/apps/cal-com) booking can create a lead in your CRM, and the [Granola](/apps/granola) note from that call can update the record afterward.
### New triggers
- **[Shopify](/apps/shopify):** new, paid, fulfilled, cancelled, and refunded orders; draft orders; new customers; new and updated products.
- **Cal.com:** new and cancelled bookings, filterable by event type.
- **Granola:** new meeting notes. **[Gong](/apps/gong):** new calls and users. **Fireflies:** new transcripts.
- **[Zoho CRM](/apps/zoho-crm):** new leads, contacts, companies, tasks, and users, plus updated leads and contacts.
### Sturdier polling
- Every trigger filter accepts multiple values, so one Gmail trigger can watch several labels.
- A Triggers tab shows each check's status and item counts, with a Sync Now button.
- Webhook triggers can authenticate with a secret URL alone, so tools that only accept a URL can start workflows.
## Multiple workspaces on one account, with readable URLs
> One login can now join and switch between several workspaces, each with its own workflows, connections, role, and default models.
Mar 29, 2026. Topics: Collaboration, Models. https://www.generalinput.com/changelog/2026-03-29-multiple-workspaces
Until now, each account belonged to exactly one workspace, so accepting a second invitation failed. As of March 29, one login can be a member of any number of workspaces. That matters most for consultants and agencies who build automations for several organizations, and for anyone who wants a test workspace next to their team's.
### How it works
- **Memberships.** Each membership links a person to a workspace with its own role and status. Workflows and connections belong to the membership, so what you build in one workspace never shows up in another.
- **Switching.** A workspace switcher sits in the sidebar, and the dashboard opens on the workspace you used most recently.
- **Readable URLs.** Dashboard links carry a workspace name you can edit instead of an internal ID.
- **Default models per workspace.** Each membership sets its own default model for chat and for building. The system default moved to Claude Sonnet 4.6.
### Also shipped
- Chats get automatic short titles, unread markers, and live run status.
- Consecutive tool calls in a chat collapse into expandable groups.
## One-click OAuth, plus Zoho CRM, Dropbox, Airtable, and more
> General Input can now supply the OAuth app so members connect in one click, starting with QuickBooks. Seven new integrations joined.
Mar 26, 2026. Topics: Integrations, Security. https://www.generalinput.com/changelog/2026-03-26-managed-oauth-and-new-integrations
At launch, every OAuth integration required a workspace admin to register an app with the provider and paste in its credentials. That's still right for teams that want tokens issued to an app they own, but it's a lot of setup for one connection. As of March 26, General Input can supply the OAuth app itself, so members connect in one click on the provider's own consent screen, starting with [QuickBooks](/apps/quickbooks).
### How it works
- **Three tiers.** Every OAuth connection uses, in order: the member's own registered app, the workspace's app, then a General Input managed app. An app you bring always wins over ours.
- **Encrypted.** Managed client secrets are stored encrypted, like every other credential.
- **Setup guides as PDFs.** Every bring-your-own setup page can be downloaded as a branded PDF for whoever administers that system.
### New integrations
[Zoho CRM](/apps/zoho-crm), [Dropbox](/apps/dropbox), [Box](/apps/box), [Airtable](/apps/airtable), [Calendly](/apps/calendly), [Cal.com](/apps/cal-com), and QuickBooks, each reading your account's structure (folders, bases, event types) as soon as you connect. [Shopify](/apps/shopify) moved from an API key to OAuth.
## Built-in workspace databases for workflows and AI agents
> Every workspace can create hosted SQLite databases that workflows, agents, and members share, with a table browser and CSV export.
Mar 26, 2026. Topics: Workflows, Agents, Platform. https://www.generalinput.com/changelog/2026-03-26-workspace-databases
Workflows often need to remember things: which leads were already processed, when the last run happened, a running list someone wants to query later. Keeping that state used to mean wiring up a spreadsheet or an outside database. Every workspace can now create its own databases, and the Databases page is in the sidebar for all members.
### How it works
- **Hosted SQLite.** Each database belongs to the workspace rather than one workflow, so one workflow can write rows another reads.
- **Agent tools.** Agents can list, create, inspect, and query databases. They reuse an existing database before creating a new one, check the schema before querying, and add a unique ID column when a workflow must never process the same item twice.
- **When agents use it.** Ask an agent to "track," "keep a list of," or "remember" something, and it reaches for a database and links you to the table it wrote.
- **Browse and export.** View tables in a data grid and export any table to CSV.
- **Isolated by workspace.** Every query is checked against the run's workspace.
- **Metered.** Queries from workflow runs cost $0.000013 each.
## Integration API docs split per endpoint for sharper agents
> Agents now search integration docs one operation at a time and load only the endpoints a task needs, instead of a service's entire API reference.
Mar 25, 2026. Topics: Agents, Integrations. https://www.generalinput.com/changelog/2026-03-25-api-docs-per-endpoint
Until March 25, when the workflow builder or the chat agent needed a service, it loaded that service's complete API reference into its context, however few endpoints the task used. A large reference crowded out the conversation and made mistakes more likely. We split every integration's documentation into individual operations, one per endpoint, plus one shared section per service covering authentication, pagination, and error handling.
### How it works
- **Search, then fetch.** Agents search across all operations with hybrid retrieval (vector embeddings plus full-text) and get the top 10 matches grouped by service, then fetch only the operations they need.
- **Everything needed to call it.** Each fetched operation arrives with its service's shared section and the exact credential fields its code should read.
- **No more whole documents.** The path that loaded an entire reference was removed, so context holds only the operations an agent looked up.
- **Shorter prompts.** Moving reference material behind search tools also shortened the agents' standing instructions, which leaves more room for your conversation and your data.
Browse every documented integration in the [integrations catalog](/apps).
## Agents render PDF reports and slide decks in a real browser
> Files the chat agent creates are now built as HTML and exported through headless Chrome as polished PDFs and images, guided by reference designs.
Mar 24, 2026. Topics: Agents. https://www.generalinput.com/changelog/2026-03-24-browser-rendered-documents
Ask the chat agent for a report, a slide deck, or a one-page graphic, and it now designs the file the way a web designer would: as HTML and CSS, exported through headless Chrome inside its sandbox. That gives every file a real layout engine, with grids, web fonts, proper text wrapping, and inline SVG charts, instead of hand-placed coordinates.
### How it works
- **Formats.** Reports export as A4 PDFs, slide decks as 16:9 landscape PDF pages, and infographics or flyers as full-page PNGs. Raw data still comes back as CSV.
- **Clean page breaks.** Agents are instructed to keep tables, headings, and charts from splitting across pages.
- **A library of reference designs.** Agents search a library of example PDF reports, slide decks, graphics, CSVs, and HTML emails by description, read the closest match, and borrow its spacing, typography, and structure rather than copying it.
- **One rendering path.** This replaced the coordinate-based PDF library and the separate PowerPoint, Word, and Excel generators. Native Office files came back in April with [document skills](/changelog/2026-04-23-document-skills).
## Chat with AI agents in Microsoft Teams and Slack channels
> The General Input agent now answers Microsoft Teams direct messages and Slack channel @mentions, with files in both directions.
Mar 23, 2026. Topics: Specialists, Integrations. https://www.generalinput.com/changelog/2026-03-23-microsoft-teams-and-slack-channels
You can now work with the agent where your team already talks. It answers direct messages in [Microsoft Teams](/apps/microsoft-teams), can be @mentioned in [Slack](/apps/slack) channels to start a thread, and sends and receives files in Slack, Teams, and [Telegram](/apps/telegram). From any of them it can search, run, and build workflows.
### How it works
- **Teams in your own tenant.** An admin creates an Azure bot and Teams app in their own Microsoft tenant by following a step-by-step guide, so the bot stays under your organization's control.
- **Members matched by email.** Teams and Slack users are matched to workspace members automatically, and one admin setup covers the whole workspace.
- **Files and formatting.** Attachments are stored with the conversation so the agent can read them, generated files come back as attachments, and replies are converted to each app's formatting.
- **Slack threads.** Replies in a thread continue the same conversation.
### Also shipped
- Agent runs moved to dedicated worker servers, separate from the dashboard's servers.
- Stop now halts a run in progress, and failed runs show a Retry button.
## Connect Postgres, MySQL, and SQL Server to AI agents
> Connect your own Postgres, MySQL, or SQL Server database, and AI agents query it in chat or in scheduled workflows through a logged proxy.
Mar 19, 2026. Topics: Integrations, Agents. https://www.generalinput.com/changelog/2026-03-19-connect-postgres-mysql-sql-server
Your own databases are now connections like any other integration. Connect [PostgreSQL](/apps/postgresql), [MySQL](/apps/mysql), or SQL Server, and agents can answer questions from your data in chat or use it inside scheduled workflows. When you have more than one, type @ and the connection's name to point the agent at exactly the right one.
### How it works
- **Schema introspection.** On connect, we read the database structure (tables, views, column types, primary and foreign keys), so agents query tables and joins that actually exist instead of guessing.
- **A logged proxy.** Queries go through a General Input endpoint that opens the connection, enforces a 30-second statement timeout, and caps results at 10,000 rows. Every query is recorded in Access logs against the credential and the run that made it.
- **Notes on connections.** Any connection can carry notes, like "read-only replica; orders live in public.orders," which agents read when choosing a connection.
- **Standard settings.** Host, port, database, user, password, an optional Postgres schema, and SSL. SQL Server connections encrypt by default.
## General Input launches: AI-written workflow automation
> General Input is live in public beta. Describe an automation in plain English and an AI agent plans, builds, tests, and runs it.
Mar 17, 2026. Topics: Workflows, Agents, Integrations. https://www.generalinput.com/changelog/2026-03-17-general-input-launch
General Input is live in public beta. You describe a job in plain English, and an AI agent turns it into a workflow: a real program that runs on a schedule or in response to events, presented as a document you can read and edit rather than a flowchart of nodes. Read the [launch announcement](/blog/introducing-general-input).
### What you could do on day one
- **Plan first, then build.** The builder agent writes a narrative plan with chips for each connected account and setting, waits for your approval, then writes TypeScript and test-runs it in an isolated sandbox.
- **No fake data.** The builder may not mock responses or hardcode anything user-specific; recipients and search terms become configurable inputs.
- **Triggers.** Run from the dashboard or by HTTP, on cron schedules, or from Gmail, Google Calendar, Google Forms, Google Sheets, Reddit, RSS, and Stripe.
- **A chat agent** (then called Geni, today the Chief of Staff) in the dashboard, Slack, and Telegram that runs workflows and does one-off work with your connected accounts.
- **Built-in services and models.** Web search, scraping, AI, and image generation with no extra keys, and a choice of Claude, Gemini, and MiniMax models.
- **Templates.** Dozens of ready-made workflows for sales, operations, and customer success.
## A credential access log for every connected account
> General Input logs each use of a connected account: which credential, what used it, and a plain-English reason, all visible in Access logs.
Feb 26, 2026. Topics: Security, Integrations. https://www.generalinput.com/changelog/2026-02-26-credential-access-log
You should be able to answer "what touched my Gmail account, and why?" without reading code. Since February 26, every time the platform decrypts one of your credentials it writes a record, and the Integrations page has an Access logs tab that lists them all.
### How it works
- **A reason on every credential.** Each account a workflow uses is declared with a plain-English access reason the agent writes, for example "Fetch recorded sales call transcripts from Gong."
- **A record on every access.** An entry is written when a workflow run uses a credential, when a trigger checks a connected account for new items, and when we register a webhook with a provider on your behalf.
- **Encrypted at rest.** Credentials are encrypted with AES-256-GCM and decrypted only for the run that needs them, after checking they belong to the person running it.
- **Refresh tokens stay with us.** For OAuth accounts the server refreshes tokens itself and passes only a short-lived access token into the sandbox.
### Also shipped
- Publishing a workflow as a template packages only its code. Variable values and connected accounts stay in your workspace.
## Workflow triggers: webhooks, schedules, and app events
> Workflows start from a webhook, a cron schedule, or new items in Gmail, Google Sheets, Calendar, Stripe, Reddit, or RSS, and each item fires only once.
Feb 26, 2026. Topics: Workflows, Integrations. https://www.generalinput.com/changelog/2026-02-26-webhook-schedule-and-poll-triggers
A workflow is only useful if it runs at the right moment. By February 26 there were three ways to start one, and they are still the three trigger types today: a webhook, a schedule, or a poll that watches one of your apps for new items. The first polling providers covered new emails in [Gmail](/apps/gmail), new rows in [Google Sheets](/apps/google-sheets), Google Calendar events, [Stripe](/apps/stripe) charges and subscriptions, new subreddit posts, and RSS feeds.
### How it works
- **Webhooks.** Any workflow can be started by an authenticated HTTP request, and the dashboard's run form is generated from the workflow's input schema.
- **Schedules.** Cron expressions with a per-workflow time zone.
- **Shared poll sources.** Each source, such as an inbox or a spreadsheet, is checked once a minute no matter how many workflows watch it, and every item is recorded so it never fires twice.
- **Broken workflows stop themselves.** If a workflow's last three runs fail, it is turned off automatically and you are told why.
### Also shipped
- A Data captured tab shows exactly what each trigger received.
- A workflow can have several triggers, each enabled on its own.
## Built-in AI, web search, and scraping on one prepaid balance
> Workflows can call an AI model, web and image search, and a scraping proxy with no extra accounts, all paid from one prepaid workspace balance.
Feb 23, 2026. Topics: Pricing, Integrations, Platform. https://www.generalinput.com/changelog/2026-02-23-platform-services-and-prepaid-credits
Most automations need an AI model, a search engine, or a way to read web pages, and none of that should require signing up with another vendor first. Since February, workflows can use built-in platform services the moment they declare them, and all usage is paid from one prepaid credit balance you top up by card.
### How it works
- **The services.** An LLM with low, medium, and high intelligence levels, internet search, image search, web scraping that returns clean text or HTML (with JavaScript rendering for sites that block plain requests), and image generation.
- **Metered precisely.** Model tokens, sandbox time, and every platform call are recorded to the millionth of a dollar, and every run shows exactly what it cost.
- **Automatic top-up.** Set a threshold and an amount, and your saved card is charged when the balance drops below it. You get a notification each time it refills.
- **Stops before it starts.** With an empty balance, new runs, retries, and agent messages are refused up front rather than failing partway through.
### Also shipped
- Workspace members: invite people by email as an Owner, Admin, or Member.
## Integrations as curated API docs, not prebuilt connectors
> General Input integrations pair a secure connection with a curated API reference the agent codes against. Over 90 services were documented by February.
Feb 19, 2026. Topics: Integrations, Agents. https://www.generalinput.com/changelog/2026-02-19-integrations-documented-for-the-agent
Most automation tools can only do what their prebuilt connectors expose. In General Input, an integration is two things: a way to connect your account (OAuth or an API key) and a curated API reference the agent reads before writing code against that service. That means the agent can use almost anything the service's API supports. By February 19, more than 90 services were documented and close to 60 could be connected, including Google Workspace, Microsoft 365, [Slack](/apps/slack), [GitHub](/apps/github), [Notion](/apps/notion), [Stripe](/apps/stripe), and [Zendesk](/apps/zendesk).
### How it works
- **One record per service.** Each catalog entry holds the connection settings, a key setup guide, and the API reference, so the connection screens and the agent's docs never disagree.
- **Hybrid search.** The agent finds the right docs by meaning and by keyword, blending embedding similarity with full-text search.
- **No guessing.** The agent must read a service's docs before calling it, and if a service isn't documented it says so and suggests a documented alternative instead of inventing endpoints.
- **Docs that keep up.** Documentation agents research new services and audit existing docs against the live API.
## An AI agent builds, tests, and repairs your workflows
> Describe a workflow and an AI agent asks questions, writes the code, tests it against your real accounts, and repairs failed runs from their logs.
Feb 10, 2026. Topics: Agents, Workflows. https://www.generalinput.com/changelog/2026-02-10-agent-builds-tests-and-fixes-workflows
In General Input you don't write workflows; an AI agent does. By February 10 the agent could ask you structured questions mid-build, write the workflow, test it in the same sandbox production uses, and repair a failed run starting from its actual error.
### How it works
- **Questions as forms.** Instead of guessing, the agent asks clarifying questions, confirmations, and connection requests as widgets in chat, then resumes exactly where it stopped once you answer.
- **Tested against real services.** Test runs use your connected accounts, and the agent must read the output before it reports back. Its instructions forbid mock data and placeholder fallbacks that make a run look successful without doing the job.
- **Fix with AI.** Any failed execution has a Fix with AI action that opens a new conversation seeded with that run's state and logs.
- **Long conversations keep going.** When a conversation nears the model's context limit, it is summarized into the request, current state, pending work, and key decisions, and the agent continues from there.
### Also shipped
- Every agent edit saves a version you can restore.
- A model picker in chat, and a running total of what building a workflow has cost.
## Workflows run as real code in isolated cloud sandboxes
> From week one, General Input workflows were TypeScript programs written by AI, each run in a fresh, isolated cloud sandbox.
Jan 29, 2026. Topics: Workflows, Platform. https://www.generalinput.com/changelog/2026-01-29-workflows-run-as-code-in-sandboxes
We started General Input on January 24 with one thesis: instead of hand-maintaining hundreds of connectors that fall behind API changes, ship identity, triggers, and execution, and let an AI agent write the integration code each workflow needs. By January 29, every workflow was a real program running in its own fresh, isolated cloud sandbox.
### How it works
- **A contract plus code.** Each workflow pairs a short configuration the platform can read (steps, inputs, and the routes between them) with TypeScript the agent writes. The dashboard draws the workflow from the configuration without reading any code.
- **A fresh sandbox per run.** Every execution gets a new E2B cloud sandbox: the code is uploaded, dependencies install, and the run streams its progress live to the dashboard.
- **Files between steps.** Steps pass data as typed files, and the final files are the run's output, so results are easy to inspect and download.
- **Secrets only at run time.** A workflow records which connected account each step needs, never the secret. Credentials are decrypted for that one execution and passed into the sandbox.
The file layout has changed since, but the split between a plan the platform can read and code the agent writes has held from the first week.