See which customers cost more in sandbox compute than they pay

Build me an app that shows which of my customers cost more in E2B sandbox compute than they pay me. I run a multi-tenant AI product and I tag every sandbox at creation with the customer it belongs to, so that metadata tag is the join key between my infrastructure bill and my revenue.

Give the app a settings page where I set the numbers everything else depends on. I need a per-vCPU-hour rate and a per-GiB-hour rate that I type in myself, defaulting to E2B's published $0.0504 per vCPU-hour and $0.0162 per GiB-hour but fully editable. I also need to choose which sandbox metadata key holds the customer identifier, because every team tags differently: mine might be customerId, someone else's is userId or orgId. Finally let me point at the Google Sheet and tab that holds my plan prices, choose the sheet the monthly ledger gets appended to, and pick the Slack channel for month-end posts. Persist all of this.

The main screen is a table of customers. Build it by calling E2B List Sandboxes with a metadata filter on the configured key, then grouping the results by that key's value. Each row shows the customer identifier, how many sandboxes they have, their total compute hours, the estimated compute cost, their monthly plan price, and the resulting margin in both dollars and percent. Read the plan prices from my pricing tab with Google Sheets Get Values and match them to the customer identifier. Let me sort by thinnest margin so whoever is underwater floats to the top, and style negative-margin rows so they are obvious at a glance.

E2B does not return dollar figures anywhere in its API, so cost has to be derived rather than fetched. For each sandbox, take its CPU and memory allocation and multiply by how long it actually ran: vCPU count times hours times my per-vCPU-hour rate, plus memory in GiB times hours times my per-GiB-hour rate. Get the allocation from Get Sandbox, and the CPU, memory and disk numbers from List Sandboxes Metrics for the batch view and Get Sandbox Metrics for a single sandbox. Work out real lifespan from List Team Sandbox Lifecycle Events by pairing each created event with its matching paused, resumed or killed event, so a sandbox that was up for twenty minutes is not billed as if it ran all week.

Clicking a customer opens their sandbox history in detail: every sandbox they have ever had, with template, current state, created and ended times, lifespan, peak and average CPU and memory, disk usage, and that sandbox's share of their bill. Show the lifecycle event history too, so I can see the actual pattern of starts, pauses and resumes.

Put an "Explain this account" button on every customer row and on their detail page. It kicks off a background agent that reads that customer's sandboxes, metrics and lifecycle events, then writes a plain-English explanation of what actually drove their bill, covering things like long-running sessions, oversized templates, or sandboxes left paused for weeks, along with concrete recommendations. Save the write-up back into the app so it shows on the customer's page with the date it was generated, and show progress while it runs.

Give me two finish-of-month actions. One appends the whole margin ledger to a Google Sheet using Append Values, one row per customer with sandbox count, compute hours, estimated cost, plan price and margin. The other posts the list of underwater accounts to Slack with Send a Message, formatted so I can read it on my phone.

One cost nuance worth baking in: paused sandboxes stop billing compute, but their filesystem and memory snapshot keeps consuming storage until they are deleted, and E2B retains them for up to thirty days. A customer sitting on a pile of forgotten paused sandboxes is a real margin leak, so surface paused sandbox counts and their age on the customer page, and have the agent call it out when it applies.