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, via Token Tunnel), 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
| General Input | Hermes Agent | |
|---|---|---|
| Self-hosted on your own infrastructureHermes: free, DIY, MIT-licensed. General Input: managed on-prem for enterprises. | ||
| Runs autonomously on schedules | ||
| Local / open-weights model supportHermes via Ollama. General Input via Token Tunnel and managed open-weights models. | ||
| Reusable skillsGeneral Input skills are authored, versioned, and shared under permissions. Hermes writes and rewrites its own. | ||
| Deterministic, reproducible runsA self-modifying agent cannot guarantee the same behavior twice. | ||
| Built for non-technical users | ||
| Human approval gates before sensitive actions | ||
| Complete, exportable audit log of every run | ||
| Least-privilege credential scopingAgent Firewall limits each credential to specific operations, enforced outside the model. Hermes can read whatever its environment holds. | ||
| Role-based access control for teams | ||
| Per-run cost transparency | ||
| Managed integration catalog500+ apps with 30,000+ documented operations and managed OAuth. |
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 teammates 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: Token Tunnel routes your existing AI subscriptions, or local models on your own hardware, into cloud workflows, so prompts run on compute you control; 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. 500+ managed integrations with OAuth handled, versus building connectivity skill by skill.
FAQs
Does General Input support local models like Hermes does?
Yes. Token Tunnel connects local models running on your own hardware, and 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.