MergeJudge documentation

Guides for the open-source AI pull request reviewer that learns your repo's rules.

Start

• Getting started: install the GitHub App and get a first review.
• How it works: rules pick a lane, the AI reads the diff, the verdict lands on the PR.
• Review modes: diff, context and deep.
• Lanes and rules: route each PR to auto, review or block.

Run MergeJudge

• Hosted service (your own key): the GitHub App, with your own model key.
• GitHub Action: run the review inside your own GitHub Actions runner.
• Run reviews on your own machine: the open-source runner.
• Self-hosted GitHub Action: a full workflow with your own AI key, deep mode and learning.
• Local AI review on your machine: the AI step on your own computer, and early feedback from draft PRs.
• Self-host on Cloudflare: current status.

Reference

• .mergejudge.yml: every option and its default.
• LLM providers: Anthropic, OpenAI, z.ai GLM, OpenAI-compatible APIs and Workers AI.
• Rubric packs: security, correctness and tests.
• Policy files: REVIEW_POLICY.md, AGENTS.md and friends.
• AGENTS.md and policy files for AI review: how policy files, .mergejudge.yml and learning fit together.
• Approvals: when MergeJudge approves, and how it works with branch protection.
• Methodology: our small offline test of the review modes.

Operate

• Learning loop: how the reviewer learns from your feedback.
• Security model: permissions, key handling and base-branch-only config.