Minimal asynchronous PR-review workflow over GitHub for AI agents (Claude Desktop, Codex, pi.dev). One TypeScript package exposes a CLI (primary) and an MCP server (secondary) over a shared core. GitHub is the source of truth: the ai-review label + a native review request route a PR to an engineer's agent, a claim-marker comment pins the reviewed commit SHA, and completion posts a native PR review.
# ~/.npmrc: @input-output-hk:registry=https://npm.pkg.github.com (+ read:packages token)
npm i -g @input-output-hk/agent-reviewagent-review labels bootstrap --repo input-output-hk/some-repo
agent-review request --repo input-output-hk/some-repo --pr 42 --reviewers yshyn-iohk --skills security,api
agent-review list --repo input-output-hk/some-repo
agent-review claim --repo input-output-hk/some-repo --pr 42
agent-review complete --repo input-output-hk/some-repo --pr 42 --event comment --summary "LGTM"MCP hosts: { "command": "npx", "args": ["-y", "@input-output-hk/agent-review", "serve"] }.
Multiple requested reviewers now run as a concurrent panel: the earliest to claim is the anchor and posts the primary review, and every other claimant is an enricher that adds one consolidated second opinion once the primary lands. See Panel review (multiple reviewers) for the full flow.
Full documentation: https://input-output-hk.github.io/agent-peer-review/
Repo setup (one-time): Settings → Pages → Source → GitHub Actions to enable the docs site.
npm install && npm test && npm run buildArchitecture decisions are recorded as ADRs under docs/adr/ (rendered in the docs site under "Architecture Decisions").
Apache-2.0.