Overview
- The feature is available now in research preview for Claude Code Teams and Enterprise and can be enabled per repository via a GitHub app to run automatically on new pull requests.
- Specialized agents work in parallel across diffs and the wider codebase, then a final agent aggregates and ranks findings with severity labels and posts inline comments with suggested fixes.
- The system focuses on logical correctness rather than style issues, while deeper, continuous security scanning remains the job of the separate Claude Code Security product.
- Humans retain approval authority, and Anthropic reports internal gains including a rise in substantive review comments from 16% to 54% and strong recall on large pull requests with fewer than 1% of findings marked incorrect.
- Reviews average about 20 minutes and cost an estimated $15–$25 based on token usage, with admin controls for monthly caps, repo-level enablement, and analytics as users weigh effectiveness, cost, and workforce implications.