How AI review works

An AI review reads a pull request the way a careful teammate would: the change, the code around it and, when there is one, the ticket it delivers. It then gives a verdict, a short summary and a list of findings, each tied to a file and a line where it can be. The review runs on your computer, with your own Claude or Codex sign-in. Limn charges nothing for it.

#Where it lives

Open any pull request and choose the AI Review tab. Everything about AI review for that pull request is on this one tab: the run controls, the latest review, earlier reviews, the review threads, the CI check, the Story check, the design diagram and the Review chat. The AI review column on Open pull requests and the AI reviews tab on Pending show the same reviews across the workspace.

#What a review does

When a review starts, Limn first decides how much reading the change needs. You do not choose this.

• Deep review: the reviewer has a copy of the repository at the pull request's latest commit, so it can open other files, follow callers and check how the change fits. On a deep review it may also ask up to three specialists for a second look: design, tests, impact on other code, security, performance and, when the change touches screens, accessibility.
• Quick review: for a small change that touches nothing other code depends on, the reviewer reads the diff alone. It is faster and uses less of your subscription.
• Skipped: when there is no change to the code to read, no model reads anything. On a review you start, the open review threads are still checked.

The reviewer can read code. It cannot change files, run commands or reach the web. Anything it reads from the pull request (the description, comments, the code) is treated as text to judge, never as instructions to follow.

#What you get

• A verdict: Approve, Comment or Request changes. It is a suggestion until a review is posted. See Verdicts and posting to GitHub.
• A summary: one plain sentence, then a bullet for each main issue, most serious first.
• Findings, each with a severity: Blocker (must change before merge), Warning (should change), Nit (small and optional) and Question (the reviewer asks the author something). A finding that cannot be placed on a line becomes a comment on the whole pull request.
• A follow-up of earlier findings when the pull request was reviewed before. See Follow-up reviews.
• A judgement of other people's threads: whether each comment holds up and whether the code now deals with it. See Check threads.

Beside the review, and as separate runs, Limn can explain failing CI (CI review), check the pull request against its ticket (Story checks) and draw how data moves through the change (Design diagrams).

#Claude, Codex or both

Limn reviews with Claude, with Codex, or with both at once. With both signed in, the default is Claude + Codex: the two review the same commit without seeing each other, and Limn merges what they found into one review. See Claude and Codex together. To sign in, see Claude and Codex.

#Nothing posts until you say so

A review you start with Review only stays in Limn. You read it, change the wording, set findings aside and post when you are ready. A review you start with Review and post goes to GitHub, under your account, when it finishes.

#What it costs

Every review runs on your own Claude Code login or Anthropic API key, or on your own ChatGPT sign-in for Codex. Limn charges nothing and stores no key. While a review runs, the tab shows the tokens used so far and, on an API key, an estimate in US dollars. On a subscription, Limn shows your share of the week instead. Automatic runs stop when your spend budget is used up. See AI usage and spend.