Overview
- The feature is live in research preview for Claude Code Teams and Enterprise customers and runs on pull-request open or update events via a GitHub app that admins enable per repository.
- Multiple specialized agents work in parallel to analyze diffs in the context of the full codebase, verify issues, rank them by severity, and leave inline comments, with humans retaining final approval.
- Admins get controls including repository scoping, monthly spending caps, and dashboards tracking reviewed pull requests, acceptance rates, and total cost.
- Reviews are token-billed and resource intensive, typically costing about $15–$25 and taking roughly 20 minutes, with time and price scaling to PR size and complexity.
- Anthropic reports internal gains such as substantive comments rising from 16% to 54%, issues found in 84% of large PRs with an average of 7.5 findings, fewer than 1% marked incorrect, and positioning that complements the separate Claude Code Security product for deeper, continuous sweeps.