PR-AF is an open-source agentic code reviewer built on the AgentField platform, and it holds the number one open-source spot on Martian’s Code-Review-Bench. Rather than summarizing diffs, it plans a deep review for every pull request and defends its findings with code evidence. The project is Apache-2.0 licensed, has about 657 stars on GitHub, and was updated in September 2026. PR-AF is a distinct product from the AgentField platform it runs on, which is listed separately on Tooloso.
For each pull request, PR-AF builds a task-specific review plan, spawns focused reviewer agents, grounds every finding in concrete code evidence and then challenges its own results before reporting. On the benchmark’s 38 runnable pull requests it reached 0.706 golden recall with the open GLM-5.2 model, ahead of dozens of compared tools including commercial reviewers, and produced 595 independently valid findings. The reviewer is model-flexible: teams can run DeepSeek-class models for routine pull requests, GLM-5.2 for deep open-model reviews, or frontier models for major changes. A one-call developer experience and one-click Railway deployment keep setup light for an agentic system.
The repository documents its architecture and benchmark methodology in the open, and the project treats reproducibility as a feature in its own right: the headline benchmark result was produced with a single open model rather than a proprietary one, so other teams can rerun the same evaluation.
PR-AF is free and open source under the Apache-2.0 license. You self-host it and pay only for the model API usage and hosting you choose.
PR-AF is for engineering teams that want a self-hosted AI pull request reviewer with benchmark-proven recall, especially teams that prefer open models and full control over their review pipeline.