AI PR Reviewer
AI code review on every pull request: comments on the exact changed lines, a summary, and the tests worth adding. A GitHub Action, a CLI and a web UI.
- Stage
- Live · open source
- Type
- GitHub Action · open source
The problem
Reviewers spend their attention on things a careful first pass would catch: an unparameterised query, a missing await, a secret in a log line. Most AI review bots either post noise or fail silently when GitHub rejects a comment.
The solution
A reviewer that comments only on real problems, on lines GitHub will accept, with suggested fixes and the tests worth adding, so human review starts from the important questions.
What's included
- — Inline comments on the exact changed lines
- — Severity levels, with optional 'changes requested' on critical findings
- — Suggested fixes you can commit from GitHub
- — Tests worth adding, in the summary
- — DeepSeek, Anthropic, OpenAI or Gemini
- — GitHub Action, npx CLI and a local web UI with review history
- — Skips drafts, lockfiles and build output; diff size budget
Built for
- — Small teams without a dedicated reviewer for every PR
- — Open-source maintainers triaging outside contributions
- — Teams that want a security pass on every change
How it works
Add one workflow file and your AI key as a secret. On every pull request the reviewer reads the diff, comments on the exact changed lines (bugs, security problems, risky patterns), writes a short summary, and lists the tests worth adding. It can request changes on critical findings, and it skips drafts, lockfiles and build output.
GitHub rejects a whole review if one comment points outside the diff, which is the most common reason AI review tools post nothing. This one sends the model real line numbers, checks every comment before posting, moves near misses to the closest changed line, and falls back to a summary if GitHub still refuses, so the findings always arrive.
Tested on a real pull request
Our demo pull request adds three routes to an orders API, with five deliberate mistakes: an injectable SQL query, an off-by-one in pagination, a missing await, an unchecked order lookup and an API key written to the logs. The reviewer flagged all five on the right lines, rated the injection and the leaked key as critical, and suggested the fixes.
Three ways to use it
As a GitHub Action on every pull request; from the terminal with npx @codeitronics/ai-pr-reviewer review <pr-link>, or on any diff before you push; or in a local web UI that shows comments inline and the history of reviews across your repositories. It works with DeepSeek, Anthropic, OpenAI or Gemini.
Open source
MIT-licensed. See a recorded review, read the code on GitHub, or install it from npm. AI review is a second pair of eyes, not a verdict: it can miss things and it can be wrong.
Screens
