What should happen?

No. A Mission can build or fix software, make a deck or a document, answer a question from your project’s sources with every claim cited, verify a pull request, or run on a schedule. It does not make phone calls, send email or messages, or generate images in this release.

An IDE agent writes code where you are typing. Boomship is the loop around a change: it reads the repository through the Brain, edits in a sandboxed runner, runs your tests, opens the result in a real browser and tries to break it, and ends with a Work Receipt that says which checks passed, failed or did not run. Many people use both.

Yours. Add Anthropic, OpenAI, Google or OpenRouter keys; they are stored encrypted and only the last four characters are ever shown again. Auto picks a model within a tier and the receipt says why. Where a deployment has managed keys, Auto can use those too.

Runs happen on the Boomship worker in an isolated runner with a private git mirror; your GitHub token never enters the sandbox. The Brain indexes a repository into rows that belong to your project. Nothing is sent to a model beyond the compiled context for the run, which the Context card shows you.

What you give it: a connected repository, files and notes you drop in, pages, and the connectors you turn on (Notion, Linear, Slack, Airtable, with Gmail and Google Drive where the host has OAuth configured). Each source shows when it was last read and whether it is in sync.

Automate chains runs and schedules them; a run that did not pass blocks everything behind it until a person looks. Signed webhooks can start a chain. One switch pauses everything automatic.

Yes. Drop files into the Brain, or work in the Playground project. Connecting a repository is one token away when you want it.

That an independent stage, not the builder, collected the diff, re-ran the tests and typecheck, had a different model review the change, scanned the added lines for secrets, and, where a page was involved, compared before and after in the browser. A check that could not run is listed as not run; confidence is rated from what actually ran.

A monthly plan for everything around the model: the runner, verification and receipts, the Brain and Automate. Inference is billed by your provider on your keys. Current plans are on the pricing page.