FAQ

Where does my code go? #

It's cloned onto Anvil's server, your server, for the length of a review and deleted afterwards. The parts Anvil needs (the diff, files it reads, the PR description and comments, CI log tails) are sent to Anthropic's API. Nothing else receives it. Details in Security and data.

Does Anthropic train on my code? #

Anthropic doesn't train models on data sent through its API by default.

What does it cost? #

There are no seats or plans. Each review, question and CI diagnosis is an Anthropic API call with a hard cost cap: $3 per review and $1 per question or diagnosis on this server. Each invited account also has a monthly AI budget (normally $25) across its repositories; when it runs out, reviews pause until the 1st. You can set a lower cap per repository with max_budget_usd. The walkthrough footer and the Agent runs page show the cost of every run. Ask the admin about anything beyond that when you get your invite.

Which code hosts does it work with? #

GitHub, through the Anvil GitHub app, and repositories hosted on Anvil itself. A GitHub repository can also be mirrored into Anvil; see Add repositories. GitLab and Bitbucket aren't supported.

How do I run my CI on Anvil? #

Change runs-on: ubuntu-latest to runs-on: anvil-4cpu in a GitHub Actions workflow, or let CI → Run on Anvil open the pull request for you. See Quickstart: CI runners. Repositories hosted on Anvil run CI here already.

Will my macOS or Windows jobs move? #

No. Anvil runs Linux x64 jobs. The migration leaves macOS, Windows, Arm and self-hosted jobs where they are; a matrix keeps its other entries on GitHub.

Where are hosted repositories stored? #

On Anvil's server, your server. See Security and data.

With a one-time link emailed to your address, or with GitHub. There are no Anvil passwords. Both lead to the same account when the address matches; see Introduction.

How do I get access? #

Request access. If it's a fit, you get an invite by email, create your Anvil account from it, and connect a repository.

How long does a review take? #

Most finish within a few minutes. Large pull requests take longer; a run stops after 20 minutes at most.

Will it review every push? #

Yes, incrementally: after the first review, each push gets a review of only the new commits. Pushes within 15 seconds of each other are reviewed together. @anvil pause stops it on one PR. See How reviews work.

Can it block a merge? #

Only through pre-merge checks in error mode, combined with a branch protection rule that requires the Anvil status. Findings on their own never block anything.

Can it change my code? #

Only when someone with write access asks. Reviews never change code. You can Apply a finding's suggestion from the workspace (the commit is yours), or ask in a comment: @anvil fix, fix ci, tests, resolve conflicts and the rest. By default Anvil puts its change on a new branch and opens a pull request into yours, so nothing lands without you merging it. See Let Anvil change code.

The review is wrong. What do I do? #

Reply in the PR with @anvil and explain. It checks the code again and says so if you're right. If it's a rule for the future, @anvil learn <rule> makes every later review follow it.

Which model does it use? #

Anthropic's Claude. Reviews are routed by size and risk: docs, formatting and bot PRs go to claude-haiku-5-5, small low-risk diffs to claude-sonnet-5-5, and large or risky ones (migrations, auth, CI config…) to claude-opus-5-5. Questions use claude-sonnet-5-5. A repository can change the model per tier with reviews.models, pin one model with model, or turn routing off. See Model routing.

Does it run my tests? #

Not yet. It reads the code and the CI results for the commit. Running a PR's tests in a throwaway container to back up a finding is planned.

Can my agent use it? #

Yes, through the MCP server: agents can run workflows, wait for them, read logs and fetch reviews with a personal API token.

FAQ