Skip to main content

Agent Peer Review

Turn a GitHub pull request into a task your AI agent can claim, review, and complete on its own.

Agent Peer Review is a minimal, asynchronous review workflow for AI agents such as Claude, Codex, and pi.dev, built so GitHub stays the single source of truth. No queue, no database, no scheduler to run.

Why it works this way​

GitHub-native

No external queue or database to run. A single ai-review label, GitHub's own reviewer request, a claim-marker comment, and a native pull request review carry the entire workflow end to end.

CLI and MCP, one core

One TypeScript core library drives both a scriptable agent-review CLI and an eight-tool MCP server, so the same review, self-review, and follow-up operations work from a terminal, a script, or an MCP host.

Label-selected skills

Attach a skill label such as security or api to a request, and the reviewer agent receives that specialty checklist layered on the default review, composed automatically the moment it claims the pull request. Programming languages need no label at all: the agent detects them from the pull request's changed files and loads the matching checklist on its own.

Guided, minimal setup

Run agent-review init once to authenticate, bootstrap labels, and write only the options you chose. Your GitHub login is auto-detected from the token; set default repositories and peer reviewers once, then reuse them from every host.

How it works​

  1. Self-review and request. An implementing agent fixes everything found in its current-head self-review, records the successful pass, then adds the ai-review label and requests a reviewer through GitHub's own Reviewers field. A maintainer may request review on somebody else's PR directly. An optional skill label such as security attaches a specialty.
  2. Claim. The reviewer agent lists its open requests, claims one, and gets back the pull request pinned to a commit SHA plus the fully composed review instructions.
  3. Complete. The agent submits a native GitHub pull request review at that pinned commit. GitHub clears the request automatically, and the claim marker is deleted.

Continue to Quick start to install the package and wire it into a host, or read Lifecycle for the full state machine behind these three steps. Review convergence explains stable finding IDs, exact-head evidence, rereview, convergence, design escalation, and the single meaningful follow-up. How it works diagrams every flow, operation, and safety rail as the code actually implements them, with the status vocabulary checked against the source by a test.