Five ways to get a review
Anthropic’s docs list these options, as of November 2026:
| Option | What starts it | Where it runs | Who can use it |
|---|---|---|---|
| Code Review | PR activity, as set per repo, or an @claude review comment |
Anthropic’s infrastructure | Team and Enterprise organizations with GitHub repos, once an Owner turns it on |
/code-review |
You type it in a session | Your own Claude Code session | Anyone, on any plan |
Ultrareview (/code-review ultra) |
You start it from Claude Code | A cloud sandbox | Requires a claude.ai login; not on Bedrock, Google Cloud or Foundry |
| Your own CI | Your pipeline, through GitHub Actions or GitLab CI/CD | Your own runners | Anyone with an API key or cloud provider credentials |
| Your own reviewer | A subagent or skill you wrote | Wherever you run Claude Code | Anyone who writes one |
How Code Review works
A review runs in four stages: several agents read the diff and the surrounding code in parallel, a verification step drops false positives, the rest is deduplicated and ranked by severity, and the result is posted on the pull request as inline comments, a summary and a check run named Claude Code Review. By default it looks for correctness: bugs that would break production, not formatting or missing tests.
Inside a review, access to the repository is read-only and it does not push commits. It reads code; it does not build or run it.
Setup is done by an Owner of the Claude organization: install the Claude GitHub App, choose repositories, and pick per repo when reviews run:
- Once after PR creation: when a PR is opened, reopened or marked ready for review.
- After every push: the default for a newly added repo, and the most reviews.
- Manual: only on an
@claude reviewcomment.
Drafts and pull requests from forks are skipped automatically. The Owner also sets a review level: Quick, Standard (the default) or Deep. Deeper costs more. You cannot pick the model.
What it costs
There is no flat price per review. Code Review is billed on token usage through usage credits, separate from your plan’s included usage, and billed to the organization rather than the developer. Cost per review grows with the review level, the size of the change, the complexity of the surrounding code and how many candidate issues need checking.
The docs suggest a pilot before rolling out:
- Set a monthly spend limit for the Claude Code Review service at claude.ai/admin-settings/usage.
- Turn it on for a few typical repos with “Once after PR creation” or “Manual”.
- After several reviews, read the Average cost column under Admin settings > Claude Code.
- Multiply by the number of PRs per month, and for “After every push” again by pushes per PR.
When the limit or the credit balance runs out, reviews are skipped and the PR gets a comment saying why.
/code-review in your own session
/code-review (or /review) reviews the changes on your current branch plus uncommitted work, in a background subagent with its own context, so it is not swayed by the conversation that produced the code. You can point it at a PR number (/code-review 1234), a branch range (main...my-feature) or a path. Add an effort level such as /code-review high for more coverage or low for fewer, more certain findings. It follows your CLAUDE.md but not REVIEW.md, and it counts toward your normal plan usage or API bill.
The misunderstanding: “no findings means the code is fine”
Anthropic says so on the page itself: Code Review may miss bugs and security issues, and a review with no findings does not mean the code is error-free or secure. It is one layer next to tests, CI and a human approving the merge. The best-practices docs add a warning for the other direction: a reviewer asked to find problems will always find something, so tell it to skip style, or you end up fixing things that did not need fixing.
What this means for a team
The review is only as good as what it knows about the team’s standards. CLAUDE.md and REVIEW.md on the default branch are where those go, and someone has to keep them current. A decision made in a meeting or a ticket reaches the reviewer only if it was written there. See Claude Code memory for how those files load, and how to prevent AI mistakes and review its work.
Where Bombos fits
A reviewer flags what breaks the rules it can see. Bombos is a plugin that gives Claude Code and the team’s other AI tools one shared memory with the team’s knowledge and the rules it learned, read also from mail, documents and repos, so those rules do not depend on someone copying them into a REVIEW.md. Every answer is traceable to its source. One developer can try it on their own agents first.
Related questions
Can Code Review block a merge?
No. A review never approves or blocks a pull request. It posts a check run named Claude Code Review, and that check run does not conclude as a failure. The merge decision stays with your team.
Does Code Review read our CLAUDE.md?
Yes, and it also reads a REVIEW.md for review-only rules. REVIEW.md is read from the base or default branch, never from the pull request's own changes, so a PR cannot rewrite its own review rules.
How do I request a review by hand?
Comment @claude review on the pull request. It runs one review. @claude review always also subscribes the PR to a review on every following push. Only people with write access can trigger it.
Is Code Review available with Zero Data Retention?
No. It is not available to organizations with Zero Data Retention or the HIPAA configuration, and not yet for organizations that use customer-managed encryption keys.
Does /code-review fix the problems too?
With --fix it applies the findings to your working tree after the review, through your normal permission mode. With --comment it posts the findings on a GitHub pull request or GitLab merge request.
