How to Set Up Claude Code Review on GitHub: Turn It On, Pick a Review Level, and Keep the Cost in Check
Image: Anthropic (official Claude Code docs) — how a Code Review run moves from a GitHub pull request through parallel agents, a verification step, and ranking, back to inline comments, a summary, and a check run
Anyone who has worked on a team that adopted coding agents knows the new bottleneck. Writing a change got faster; reading it did not. Pull requests pile up, reviewers skim, and the bug that matters slips through in a 900-line diff. Anthropic's answer is Code Review for Claude Code, a managed service that reads every pull request in the GitHub repositories you choose and leaves inline comments before a human gets there. On October 15, 2026 it left research preview and became generally available to Claude Team and Enterprise organizations.
This guide walks through what you need, how an organization Owner turns it on, how developers use it day to day, and the settings that decide how much it costs. Every step comes from Anthropic's own announcement and the Claude Code documentation.
What it actually does
When a review runs, several agents examine the diff and the code around it in parallel. A verification step then checks each candidate finding against the code and drops false positives. What survives is deduplicated, ranked by severity, and posted back to the pull request as inline comments, a summary in the review body, and a check run named Claude Code Review.
Two design choices are worth knowing up front. First, a review never approves or blocks a pull request, and its check run never concludes as a failure, so merging stays a human decision. Second, it works by reading code, not running it. Inside a review, repository access is read-only and it does not push commits. By default it focuses on correctness, meaning bugs that would break production, rather than formatting or missing tests.
Anthropic says the GA version is faster than the preview, with most reviews finishing in about 6 to 10 minutes instead of roughly 20, and that admins can now pick a review depth per repository.
Check that you qualify
- Plan: Claude Team or Enterprise. If you're on another plan, skip to the
/code-reviewsection below. - Code host: GitHub, including GitHub Enterprise Server and GitHub Enterprise Cloud with data residency. Repositories on other hosts aren't covered; GitLab teams can run Claude in their own CI instead.
- Data settings: it isn't available to organizations with Zero Data Retention or the HIPAA configuration, and it isn't yet available with customer-managed encryption keys.
- People: you need an Owner of the Claude organization, an owner of the GitHub organization to install the Claude GitHub App, and write access to each repository you add.
If your developers use Claude Code through Amazon Bedrock, Google Cloud, or Microsoft Foundry, you still need a Claude organization on one of these plans.
Step by step: turning it on (for Owners)
- Open the admin page. On claude.ai, go to Admin settings > Claude Code, find the Code review section, and click the Allow code review switch.
- Connect GitHub. If a Connect GitHub dialog appears, click Connect and authorize Claude. GitHub Enterprise Server users pick their instance first.
- Install the Claude GitHub App. If your repositories don't show up in the Add a repository dialog, click Install GitHub App, pick the GitHub organization, choose which repositories the app can access, and approve the permissions. Note that the app's permission set is broader than a review uses, because other Claude features share it.
- Choose repositories and click Continue.
- Click Finish on the "Code review is set up" dialog. If you close it any other way, your repositories stay added but the main switch stays off.
- Test it. Open a non-draft pull request in one of those repositories. A Claude Code Review check run should appear within a few minutes.
The two settings that matter most
Review level. There are three: Quick (fastest and cheapest), Standard (the default), and Deep (most thorough, slowest, most expensive). You set one organization default and can override it per repository. A sensible pattern is Quick for internal tools and docs, Deep for payments, auth, or anything safety-critical.
When reviews run. Each repository has a Review runs setting with three options:
- Once after PR creation: one review when a pull request is opened, reopened, or marked ready.
- After every push: a fresh review on each push, which also auto-resolves threads once you fix them.
- Manual (@claude review): nothing runs until someone asks.
Here is the trap: a newly added repository starts at After every push, the option that runs the most reviews. If your team pushes ten times per pull request, that's ten reviews. Decide on purpose rather than leaving the default.
How developers use it on GitHub
Once it's on, developers don't configure anything. Findings appear as inline comments marked by severity: 🔴 Important (fix before merging), 🟡 Nit (minor), and 🟣 Pre-existing (a real bug the pull request didn't introduce). Many findings include a collapsed "Why this was flagged" section explaining how Claude verified the problem.
To handle a finding, push a fix (the next review resolves the thread if it judges the issue fixed) or resolve the thread yourself to dismiss it. Replying alone doesn't dismiss anything and doesn't start a review.
Comment commands give developers control without admin rights:
@claude reviewruns one full review now, which is useful in Manual repositories, on drafts, and on pull requests from forks, which automatic reviews skip.@claude review alwaysruns a review and subscribes that pull request to a review on every later push.
The command must be a new top-level comment at the start of the line, from someone with write access. Look for the check run titled "Code review queued..." to confirm it was accepted. If your repository also runs the Claude Code GitHub Action with the default @claude trigger, change the Action's trigger_phrase so the two don't both fire.
Teach it your team's rules with REVIEW.md
Code Review reads your CLAUDE.md files as project context, and a separate REVIEW.md at the repository root for review-only rules. This is where you cut noise. In practice, the most useful instructions are:
- Redefine what counts as Important for this repository.
- Cap nits, for example "Report at most five Nits per review."
- Skip generated files, lockfiles, and anything CI already enforces.
- Add checks you always want, such as "new API routes have an integration test."
One detail catches people out. REVIEW.md is read from the base or default branch, never from the pull request itself, so merge changes to it first. Some instructions, like lowering severity or excluding paths, are only honored when the default branch is protected.
Keeping the cost predictable
There's no flat per-review price. Reviews are billed on token usage through usage credits, separately from your plan's included usage, and charged to the organization rather than the developer. Cost scales with review level, change size, code complexity, and how often reviews run.
Anthropic's own suggested rollout is sensible. Set a monthly spend limit at claude.ai/admin-settings/usage, pilot on a few representative repositories set to Once after PR creation or Manual, then read the Average cost column in the Repositories table and multiply by your monthly pull request count. Activity and estimated spend appear at claude.ai/analytics/code-review. When a limit or credit balance runs out, reviews are skipped and a comment on the pull request says why. Anthropic also says eligible Enterprise organizations can claim at least $10,000 in Code Review credits by January 15, 2027, with terms and a 30-day expiry after claiming.
Not on Team or Enterprise? Use /code-review
On any plan, you can type /code-review inside a Claude Code session to review your current branch, or add a target such as a pull request number (/code-review 1234) or a range (/code-review main...my-feature). Add --fix to apply findings to your working tree or --comment to post them on a GitHub pull request. It runs in your own session and counts toward your normal usage. Note that it follows CLAUDE.md but ignores REVIEW.md.
My take
The interesting part isn't that an AI reads pull requests; plenty of tools do. It's the guardrails: a verification step before anything is posted, no power to block a merge, read-only access, and per-repository control over depth and frequency. That makes it a reasonable first reviewer rather than a replacement. If you're an admin, the practical move is a two-week pilot on two or three repositories, set to Once after PR creation, with a spend cap in place. Count how many Important findings were real bugs your humans would have missed. That number, not the launch claims, should decide whether you roll it out wider. And keep the documentation's own warning in mind: a review with no findings doesn't mean the code is bug-free.
Sources
- Claude by Anthropic: Code Review is now generally available (October 15, 2026)
- Claude Code Docs: Code Review
- Claude Code Docs: Set up Code Review
- Claude Code Docs: Read and request reviews on a pull request
- Claude Code Docs: Customize Code Review
- Claude Code Docs: Code Review costs and usage
- Claude Code Docs: Review locally with /code-review
댓글
댓글 쓰기