Claude Code review: managed PR reviews and /code-review
Claude Code Review on GitHub PRs (Team and Enterprise, billed by token use), /code-review and ultra in the terminal, and the GitHub Action for CI.
“Claude Code review” is two things. Code Review is Anthropic’s managed reviewer for GitHub pull requests, on Team and Enterprise plans only: an Owner turns it on in Admin settings > Claude Code, installs the Claude GitHub App, and Claude then leaves inline comments, a summary and a check run on each PR, billed by the tokens each review uses, through usage credits. It never approves or blocks a PR. On any plan, you can instead type /code-review (alias /review) in a Claude Code session to review your own branch, /code-review ultra for a deeper cloud review, or /security-review for security issues only. To review every PR in your own CI on your own key or plan, use the anthropics/claude-code-action GitHub Action.
Which one you can use
Anthropic’s Code Review documentation compares the options. As of November 2026 (Claude Code 2.1.330):
- Code Review (managed): starts on pull request activity or an
@claude reviewcomment, runs on Anthropic’s infrastructure, needs a Team or Enterprise organization with repositories on GitHub. /code-review: you run it on your branch, it runs in your own session, and it works on any plan.- Ultrareview (
/code-review ultra): you start it from Claude Code, it runs in a cloud sandbox, and it needs a claude.ai sign-in. - The GitHub Action: runs on your own CI runners on events you choose, for anyone with a Claude API key, a subscription token or cloud provider credentials.
Which plans carry Claude Code at all, and at what price, is in which Claude plans include Claude Code.
Code Review on GitHub pull requests
Who can have it. Team and Enterprise plans, with repositories on GitHub, GitHub Enterprise Server, or GitHub Enterprise Cloud with data residency. The docs say it “isn’t available to organizations with Zero Data Retention enabled or with the HIPAA configuration applied”, it is not covered by Anthropic’s BAA, and it “isn’t yet available” to organizations using customer-managed encryption keys. A team that uses Claude Code through Amazon Bedrock, Google Cloud or Microsoft Foundry still needs a Claude organization on one of these plans.
Setting it up. Per Anthropic’s setup page, an Owner of the Claude organization goes to the Code review section of Admin settings > Claude Code, clicks Allow code review, connects GitHub, installs the Claude GitHub App on the repositories (a GitHub organization owner has to do or approve this), picks repositories and clicks Finish. No workflow file goes in the repository. Open a non-draft test PR and a check run named Claude Code Review should appear within a few minutes.
Per repository, the Review runs setting is one of Once after PR creation, After every push or Manual (@claude review). A newly added repository starts at After every push, which runs the most reviews and so costs the most; set it lower before a wide rollout. The review level is Quick, Standard (the default) or Deep, set for the organization and overridable per repository. There is no model choice.
What it posts. Inline comments on the lines concerned, a summary in the review body, and the check run, whose Details list every finding by severity: 🔴 Important (fix before merging), 🟡 Nit, and 🟣 Pre-existing (a bug the PR did not introduce). The check run concludes neutral, never as a failure, so a review does not fail a required check. The review reads the diff and surrounding code without building or running anything, and its access to the repository is read-only. To dismiss a finding, resolve its thread; a reply alone does not, and a later review resolves the thread itself when a push fixes the issue.
Steering it. A REVIEW.md at the repository root holds review-only rules (what counts as Important, paths to skip, a cap on nits), and CLAUDE.md files are read as project context. Both are read from the base or default branch, never from the PR’s own changes, and exclusions and nit caps are honoured only when that branch is protected. Details are on Anthropic’s customization page.
@claude review and the July 2026 change
Anyone with write, maintain or admin permission on the repository can start a review with a comment. It has to be a new top-level PR comment (not an inline one, not an edit), with the command on one line at the start; bots cannot start one. The pull request page lists the commands:
@claude reviewor@claude review once: one review now, covering the whole PR.@claude review always: a review now, and one on every later push to this PR.@claude review stop(oroff,pause): ends that subscription.
The behaviour changed in July 2026. The docs say: “Before a July 2026 update, @claude review subscribed the PR to push-triggered reviews. If you relied on that behavior, comment @claude review always instead.” Each review a developer requests is billed to the organization.
Drafts and forks. Automatic reviews skip both, whatever the Review runs setting. On a fork’s PR, someone with write access to the base repository comments @claude review; always does not subscribe a fork’s PR to later pushes. If the repository also runs the Claude Code GitHub Action with its default @claude trigger, an @claude review comment starts the Action too; give the Action a different trigger_phrase.
What a managed review costs
There is no flat price. The costs page says reviews are “billed on token usage through usage credits, separately from your plan’s included usage”, to the organization rather than the developer, and the cost of each scales with the review level, the size of the change and the complexity of the code around it. When Anthropic announced the feature in March 2026, its post Bringing Code Review to Claude Code put reviews at an average of “$15–25, scaling with PR size and complexity”; the current docs give no figure, and instead show each repository’s real average in the Average cost column of the Repositories table. The product page says “Most reviews finish in 6 to 10 minutes.”
Capping spend: set a monthly limit for the Claude Code Review service at claude.ai/admin-settings/usage. When that limit, the organization’s overall limit or the credits balance runs out, reviews are skipped and the PR gets a comment naming the cause. Covering a whole GitHub organization at once requires a spend limit to be set first. Activity and spend are at claude.ai/analytics/code-review.
Anthropic’s own figures for its internal use: before Code Review, 16% of its PRs got substantive review comments, and 54% did after; on PRs over 1,000 changed lines, 84% got findings, averaging 7.5 issues; and “less than 1% of findings are marked incorrect”, per the announcement. These are Anthropic’s numbers on its own repositories, not an independent measurement.
Reviewing your own changes with /code-review
/code-review is a bundled command in Claude Code; /review is its alias. It looks for correctness bugs, reports in the conversation, and counts toward your plan’s usage or API billing like the rest of the session. With no target it covers your branch’s commits ahead of upstream plus uncommitted changes. From Anthropic’s page on the command:
/code-review # current branch + uncommitted
/code-review 1234 # a pull request by number
/code-review main...my-feature # one branch against another
/code-review HEAD~15..HEAD # the last 15 commits
/code-review src/auth # a path
/code-review high --fix # effort level, then apply fixes
/code-review --comment 1234 # post findings as inline PR commentsCopiedThe effort level comes first: low reports only what it is most confident in, and medium through max widen coverage. With no level typed, it reuses the last one you typed, even from an earlier session. --max-findings all lifts the usual cap on findings (Claude Code 2.1.288 or later). In a terminal the review runs as a background subagent with its own context, so you can keep working; in a non-interactive claude -p run it waits and prints the findings, which is how you would run it headless on a server. It follows your CLAUDE.md but not REVIEW.md, which only the managed service reads.
A neighbouring command, /simplify, reviews for cleanup and applies it, and does not look for bugs.
Ultrareview: /code-review ultra
/code-review ultra runs a deeper review in a cloud sandbox instead of your session. /ultrareview is an alias where it is available to your account. It is a research preview. Per the ultrareview page:
- Sign-in: needs a claude.ai account, so an API-key-only login has to run
/loginfirst. Not available on Bedrock, Google Cloud’s Agent Platform or Foundry, or with Zero Data Retention or HIPAA, where it falls back to a local review; CMEK organizations cannot run it. - Price: three free runs on Pro, Max and Ultra, once per account, then “typically $5 to $25 in usage credits”, shown in the launch dialog before you confirm. Usage credits must be turned on.
- Time and size: typically 5 to 15 minutes; by default up to 500 changed files and 8,000 changed lines.
- Targets: your branch against the default branch,
/code-review ultra developfor another base, or/code-review ultra 1234for a github.com PR, which it clones in the cloud rather than uploading your tree.
From a script or CI job, use the subcommand, which waits and prints the findings to stdout; claude -p '/code-review ultra' exits before they arrive:
claude ultrareview
claude ultrareview 1234 --post # comment on the PR as you
claude ultrareview 1234 --trust-workspace # untrusted checkout, CICopied--trust-workspace needs Claude Code 2.1.309 or later, and --timeout defaults to 45 minutes.
/security-review
The commands reference describes /security-review as analysing “the changes on your current branch for security vulnerabilities”: it reviews the diff between your branch and origin’s default branch for risks such as injection, auth issues and data exposure. It needs an origin remote; if it fails with an ambiguous argument error, the reference sends you to its error page for that case. For the same check on every PR, Anthropic publishes a separate Action, anthropics/claude-code-security-review, which its help article says runs when a PR is opened and posts inline comments.
A self-hosted review in CI with the GitHub Action
Claude Code GitHub Actions runs Claude Code on your own GitHub runners. It is the route for anyone not on Team or Enterprise, and for teams who want to choose the prompt, model and triggers themselves. The quickest setup is /install-github-app in a Claude Code session in the repository (github.com only; needs admin access to the repository and the gh CLI signed in), which installs the app, stores the secret and opens a PR with the workflows, including a review workflow if you select it.
The credential is either an ANTHROPIC_API_KEY secret, billed as API usage, or a CLAUDE_CODE_OAUTH_TOKEN made with claude setup-token, which bills runs to your Pro, Max, Ultra, Team or Enterprise subscription. Each run also uses GitHub Actions minutes. Anthropic’s review example, trimmed to its working lines:
name: Code Review
on:
pull_request:
types: [opened, synchronize, ready_for_review, reopened]
jobs:
review:
runs-on: ubuntu-latest
permissions:
contents: read
pull-requests: read
issues: read
id-token: write
steps:
- uses: actions/checkout@v6
with:
fetch-depth: 1
- uses: anthropics/claude-code-action@v1
with:
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
plugin_marketplaces: "https://github.com/anthropics/claude-plugins-official.git"
plugins: "code-review@claude-plugins-official"
prompt: "/code-review:code-review --comment ${{ github.repository }}/pull/${{ github.event.pull_request.number }}"
claude_args: '--allowedTools "mcp__github_inline_comment__create_inline_comment"'Copied--comment is what makes it post on the PR; without it the findings stay in the workflow log. Keep the claude_args line, which the docs say is what starts the tool that posts inline comments. On a subscription, swap the anthropic_api_key line for claude_code_oauth_token. On public repositories GitHub withholds secrets from runs on fork PRs, so this workflow reviews only PRs from branches in the same repository. By default the Action authenticates to GitHub as the Claude GitHub App; if you would rather run the review from your own server with a narrower GitHub credential, GitHub access for a coding agent on a server covers the options.