PR-AF
An open-source agentic code reviewer built on AgentField that dynamically generates review pipelines.
Top-ranking open-source reviewer with evidence grounding and compound vulnerability synthesis.
PR-af is an open-source code reviewer that structures pull requests into task-specific review plans and spawns focused agent threads. It suits engineering teams looking to run deep architectural and security audits as a final CI/CD gatekeeper rather than fast interactive tools.
What it is
The system dynamically morphs its execution graph based on the topology of the incoming pull request, generating semantic, mechanical, and systemic lenses. It extracts programmatic AST snippets, runs falsifiability gates to prune unsubstantiated claims, and synthesizes compound vulnerabilities across multiple files.
Who it's for
It suits teams managing complex codebases who need rigorous, evidence-grounded GitHub review comments and are willing to trade execution speed for high accuracy. It does not suit developers looking for rapid, seconds-long interactive coding assistants for inner-loop development.
Key features
Dynamic Pipeline Architecture
Structurally morphs its execution graph based on incoming pull request topology.
Evidence Grounding
Pulls exact caller snippets and import context to verify findings before reporting.
Falsifiability Gates
Actively tries to invalidate findings based on safe behavior or existing mitigations.
Compound Synthesizer
Clusters related risks across files to evaluate larger systemic issues.
Up and running in 3 steps
Our honest verdict
What's great
- Achieves top-tier open-source performance on Martian Code-Review-Bench using models like GLM-5.2.
- Drastically reduces false positives through AST extraction and evidence verification layers.
- Identifies multi-file compound vulnerabilities that standard diff-focused tools miss.
- Runs self-hosted via Docker and integrates directly with CI pipelines, CLI, or curl.
Where it falls short
- Reviews take roughly 35-50 minutes to complete, making it unsuitable for quick inner-loop checks.
- Requires maintaining an AgentField control plane alongside API keys for underlying models.
- Incur API token costs for LLMs like GLM-5.2 or frontier models during extensive reviews.
Pricing & license
The tool is free under its open-source license, but you must supply your own API keys for underlying LLMs and provide the infrastructure to run the AgentField control plane.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
How does PR-af differ from tools like Claude Code?
PR-af is built for deep CI/CD architectural audits and takes 35-50 minutes using a parallel pipeline, whereas Claude Code is designed for fast, iterative inner-loop development.
What models can I run with PR-af?
You can run cheaper models for routine PRs, GLM-5.2 for open-model CI gates, or Opus-class frontier models for high-stakes reviews.
