Claude Code Review: Multi-Agent Pull Request Reviews

Claude Code Review is a managed GitHub review service for Team and Enterprise organizations. Multiple agents analyze a pull request, verify candidate findings, and post non-blocking comments on GitHub.

Code Review is a Claude Code research preview, not a built-in Cowork review button. An organization administrator enables it, installs the Claude GitHub App, selects repositories, and decides when reviews should run.

Availability

  • Team and Enterprise subscriptions
  • GitHub repositories approved by an organization admin
  • Not available for organizations with Zero Data Retention enabled

Reviews run on Anthropic infrastructure. Anthropic says they take about 20 minutes on average, but cost and duration vary with the size and complexity of the pull request.

What happens during a review

  1. Several agents inspect the changed code and surrounding repository context.
  2. Candidate problems are checked against the actual code to reduce false positives.
  3. Duplicate findings are removed and the remainder are assigned a severity.
  4. Claude posts inline comments and a summary through a neutral GitHub check.

The service does not approve, reject, or block a pull request by itself. Human reviewers remain responsible for architecture, product intent, risk acceptance, and the merge decision.

Severity labels

LabelMeaning
ImportantA bug that should be fixed before merging
NitA smaller issue worth considering
Pre-existingA problem that was already in the codebase

There is no official promise of a sub-1% false-positive rate. Verification is designed to reduce noisy findings, but every comment still needs human review.

Set it up

  1. An admin opens the Claude Code section in Claude organization settings.
  2. Install the Claude GitHub App and grant it access only to the intended repositories.
  3. Choose a review behavior for each repository: once when a pull request opens, after every push, or manual.
  4. Open a test pull request and confirm that the Claude Code Review check appears.

Manual reviews can be requested with a top-level @claude review comment. Use @claude review once when you want a single review without subscribing the pull request to future pushes.

Make reviews fit your repository

Code Review reads project guidance from CLAUDE.md. A root REVIEW.md can set review-specific rules such as:

  • What counts as an Important issue
  • How many nits to report
  • Which generated or vendor paths to skip
  • Which repository-specific rules to check
  • What evidence a finding must include

Keep review guidance short and concrete. Existing automated checks should continue to handle formatting, types, and tests where possible.

Frequently asked questions

Is this available on Pro or Max?

No. The managed Code Review service is currently documented for Team and Enterprise subscriptions.

Does it replace human review?

No. It is a pre-review and second-opinion system. Humans should still own product intent, architecture, security decisions, and final approval.

Can it block a merge?

The built-in check is neutral and does not block the pull request. Teams can build their own policy around the machine-readable severity summary if they deliberately want a gate.

Can I run a review from Cowork instead?

You can ask Cowork or Claude Code to inspect a repository as a normal task, but that is different from the managed GitHub Code Review service described here.

Related Guides

Related Features

  • Channels — Push developer events into a Claude Code session
  • Dispatch — Assign desktop follow-up work from your phone after a review lands

Related Comparisons