AI PR Review
https://github.com/snarktank/ai-pr-review- Category
- Developer Tools
- Rank
- No. 1676Tools index
Previous survey · No. 1649 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- snarktank
- GitHub
- 61 stars
- Date
About
Robust GitHub Actions workflow that provides AI-powered code reviews using Amp or Claude Code.
What it does
AI PR Review watches pull request events, collects the description and patch, then sends the changes to either Amp or Claude Code. It turns the model response into a summary, inline comments, and a commit status on GitHub.
Why it's ranked here
The workflow covers the full automation loop and handles transient model failures with retries, timeouts, and a fallback response. The strongest reservation is failure signaling: exhausted attempts still produce a successful status, which can hide a missing analysis.
What's good
It limits GitHub permissions to repository reads plus pull request and status writes. Concurrent runs for the same pull request cancel older work. Teams can customize the prompt and choose comments, approval, or change requests without building an integration service.
Tradeoffs
Setup requires GitHub Actions, write permissions, and an Amp or Anthropic credential. Each run installs the latest selected provider CLI when absent, reducing dependency reproducibility. XML handling uses pattern matching rather than a dedicated parser, and automated testing consists of live pull requests in a separate repository.
How to use it well
It suits GitHub teams that want an automated first pass before human approval, especially on small, focused pull requests. Start with comment-only behavior, tune the prompt to local standards, and watch workflow logs and provider quotas. It does not replace deterministic linting, tests, or human merge judgment.
Technical notes+
.github/workflows/ai-pr-review.yml runs on four pull request event types under Node.js 20 and Ubuntu, with a 20-minute job timeout and per-PR concurrency cancellation. It installs @sourcegraph/amp@latest or @anthropic-ai/claude-code, gathers metadata and patches through gh, retries the model command up to five times with a 300-second timeout, validates basic XML markers, and parses fields through Node.js regular expressions. After all retries fail, it emits fallback XML whose state is success. package.json declares Node.js 20 or newer, points its main entry at the workflow, and defines a placeholder test script. AGENT.md and CONTRIBUTING.md specify live workflow testing rather than build or lint automation.
Observed
- License
- MIT License
- Primary format
- YAML GitHub Actions workflow with embedded shell and Node.js scripts
- Install surface
- Copy the workflow into a GitHub repository and add an Amp or Anthropic API key as an Actions secret
- Interface
- GitHub pull request automation that posts a summary, inline comments, and commit status
- Platform
- GitHub Actions on Ubuntu with Node.js 20
- Provider support
- Amp CLI by default, with documented Claude Code configuration
- Testing structure
- The package test command is a placeholder; contribution guidance calls for live testing in a test repository
Read from README.md, package.json, LICENSE, AGENT.md, CONTRIBUTING.md, .github/workflows/ai-pr-review.yml.
What it can do
Analyze code changes in pull requests
GitHub pull request with code changes → AI-generated code review comments
Identify potential bugs and issues in code
Source code files in pull request → List of detected issues and recommendations
Suggest code improvements and optimizations
Code changes in GitHub pull request → Improvement suggestions and optimization recommendations
Automatically trigger code reviews on pull request events
GitHub pull request creation or update → Automated AI code review workflow execution
Check code quality and adherence to best practices
Source code files → Code quality assessment and best practice violations
Generate inline comments on specific code lines
Individual lines of code in pull request → Targeted inline review comments
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.