Claude Code Review is Anthropic's managed code review service for Claude Code. On October 15, 2026, Anthropic announced it has left research preview and is generally available to Team and Enterprise organizations: once an admin enables it and installs Claude's GitHub App, new pull requests are automatically reviewed first by a fleet of agents, with findings posted back as inline comments, a summary and a check run.
How it reviews: parallel hunting, then a vote
This is not a single model scanning the code once. Several agents look for candidate problems in parallel, then a separate group verifies each finding, removes duplicates and ranks what survives by severity; every posted comment has to state what the problem is, what triggers it and how to fix it. By default it focuses on correctness issues that could affect production, not style preferences or missing tests. In a blog post on how its own team works, Anthropic said the share of its internal PRs receiving substantive review comments rose from 16% to 54% after the preview launched, that most PRs over a thousand lines get findings, and that most small ones come back clean. Whether AI review is accurate enough still has to be proven on each team's own data; the GitHub ReviewBench benchmark covered earlier on this site was built to quantify exactly those misses and false alarms.
Three things changed at general availability
First, speed: most reviews now finish in 6 to 10 minutes, roughly a third of the preview's typical time. Second, finding quality: conclusions are shorter, with the trigger and the fix stated in a few sentences. Third, control: each repository can be set to Quick, Standard or Deep review, with Standard as the default and deeper reviews costing more. Triggers are configurable too: on PR creation, after each push, or only when someone comments claude review; draft PRs and forks are excluded from automatic review. A repository's review standards can be written into REVIEW.md and CLAUDE.md so the service follows the team's own rules.
What it does not do
The service has read-only access to repositories. It does not build or run code, and it never approves or blocks a merge; that decision stays with humans. A review with no findings does not mean the code is free of bugs or security issues. Pricing follows review depth, and eligible Enterprise organizations can claim at least $10,000 in review credits by January 15, 2027, with credits expiring 30 days after they are claimed. As coding agents push output to its limit, review becomes the narrowest stretch of the pipeline; by turning review into a managed service with a speed dial per repository, Anthropic is selling capacity for exactly that bottleneck.
