assign the issue.
come back to a PR.

Yes, and there's no card required to start. On the free tier you get 5 tasks a day when you bring your own model key (or connect a Codex / ChatGPT subscription), or 1 task a day on our keys. That's enough to watch Sinatra take a real issue from assignment to an open PR on your own codebase. When you're ready for a team, the Monthly plan is a flat $20 per member per month: hosted sandboxes and orchestration, with no markup on your LLM bill (you pay model cost at provider rates, or keep using your own key). Enterprise adds a custom annual contract. You can move from free to paid whenever it makes sense, and nothing about the workflow changes.

If you haven't connected anything, Sinatra runs on our keys and passes the model cost through at cost, with no markup. Connect an Anthropic or OpenRouter API key, or a Codex / ChatGPT or Claude subscription, and it uses yours instead. With nothing pinned it picks based on what you have connected, then posts the engine, model and credential it chose straight into the ticket, so a run is never a black box. Pin a specific model per repository in .sinatrarc when you want predictability, and pin a separate one for the planning and brainstorm phases if you'd rather those ran on something cheaper. Pin a model you have no credential for and the run stops at startup naming the missing credential, before it spends anything. Free-tier tasks that run on our keys use a single fixed model.

No. Sinatra only ever works in branches and opens pull requests. It never merges, deploys, or runs anything against your production environment. Merging stays with you and your CI, exactly like a human teammate, so your usual branch protection and required reviews still gate everything. Each task runs in its own isolated, short-lived sandbox: Sinatra checks out your repository, does the work there, and shuts the sandbox down when the task is done; idle sandboxes are archived automatically within a day. Any credentials or model keys you connect are stored encrypted and used only to run the task you assigned. And we don't use your code or content to train our own models, and we don't sell it.

In an isolated sandbox that only your workspace can use. Sinatra checks out your repository there, does the work, and stops the sandbox when the task ends; a sandbox left idle stops after two hours and is archived after a day. Sandboxes are never shared between workspaces: reusing one requires either a matching workspace label or a workspace-scoped name, and anything that can't be proven yours is treated as not found. We never copy your checkout into our database. We do keep repository metadata, the branch and pull request we opened, the commit we started from, per-run token and cost totals, and the raw webhook payloads GitHub and Linear send us, which carry issue and comment text and, on a review comment, the diff hunk it was left on. The sandbox can't reach the open internet either, only a default-deny allowlist covering GitHub, Linear, your model provider and the usual package registries.

It does its best work on big codebases. Each task runs in its own sandbox with the full repo checked out, so the agent reads and searches across the whole codebase before it touches anything. For very large repos, .sinatrarc can switch the checkout to a blobless clone so the initial clone stays fast.

It's built to. The agent is told to add or update tests for behavior changes, and to read your existing tests first so the new ones match your conventions rather than inventing a second style. Every run finishes with a written record of what it actually verified and what risk it knowingly left behind. Point .sinatrarc at your test command and Sinatra runs it in the sandbox once the agent is done, re-running the agent to fix failures up to twice. Tests that still fail don't block anything: the pull request stays open with the failure reported on the ticket, because a red test you can see beats a run that quietly gave up. Sinatra also reviews its own diff before marking the PR ready, and thin assertions and uncovered branches are among the things it looks for.