Open-Source AI Code Reviewer PR-AF Tops Benchmark, Outperforms CodeRabbit and GitHub Copilot at a Fraction of the Cost
Summary
Open-source AI code reviewer PR-AF claims the #1 spot on Martian Code-Review-Bench with a golden recall score of 0.706, outperforming commercial giants like CodeRabbit, GitHub Copilot, and Greptile while costing roughly 10 times less per review.
Key Points
- PR-AF is the #1 open-source code reviewer on Martian Code-Review-Bench, achieving a golden recall score of 0.706 using the GLM-5.2 model across 42 compared tools, surpassing all commercial competitors including CodeRabbit, Greptile, and Copilot variants.
- Unlike static review tools, PR-AF dynamically compiles review dimensions from each pull request, spawns specialized ephemeral reviewer agents, and runs evidence grounding, compound vulnerability synthesis, and falsifiability gates to eliminate false positives before posting inline GitHub comments.
- PR-AF is deployable via Docker Compose, Railway, or GitHub Actions, supports model flexibility through OpenRouter, costs approximately 10x less per review than closed-source tools, and is built on the AgentField control plane ecosystem alongside companion agents SWE-AF and SEC-AF.