PR Agent is an open-source code review agent that acts on a pull request through commands posted as comments or run from a terminal. /describe generates a title, summary, walkthrough, and labels; /review returns findings covering security, review effort, and tests; /improve returns actionable code-improvement suggestions; and /ask answers free-text questions about the change.
Each tool uses a single model call, taking roughly 30 seconds, and a PR compression strategy fits both small and large diffs into the prompt. Prompting is JSON-based, so review categories and behaviour can be changed through configuration files. Supported git providers are GitHub, GitLab, Bitbucket, Azure DevOps, and Gitea, and it can be deployed as a GitHub Action, a webhook app, a Docker container, or a local command-line tool.
Model choice is open: OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek, Mistral, and any model reachable through LiteLLM, including Azure OpenAI, AWS Bedrock, Vertex AI, Databricks, OpenRouter, and Ollama.
Version 0.39.0 feeds AGENTS.md and similar repository context files to /review, /describe, and /improve by default, and adds Agent Skills (SKILL.md), organization-level settings, a restricted mode for reduced GitHub permissions, and GitHub Checks as an output target. Version 0.40.0 moves the default model to GPT-5.6, adds Gemini 3.6 support, persistent inline comments that avoid duplicate suggestions across runs, an OpenRouter provider-routing and reasoning config, a tokenless plain-diff provider, and CI artifact context injection. Version 0.41.0 adds Claude Opus 5 support. The /help_docs tool is disabled from version 0.36.1 onward, pending a fix for a credential-exposure issue.
Features
- Comment-driven tools: /describe, /review, /improve, and /ask run as pull-request comments or from the command line
- Single call per tool: each tool issues one model call, taking about 30 seconds
- PR compression: a compression strategy fits diffs of any size into a manageable prompt
- Git provider coverage: GitHub, GitLab, Bitbucket, Azure DevOps, and Gitea
- Deployment options: command line, GitHub Action, Docker, self-hosted webhooks
- Model flexibility: OpenAI, Anthropic, Google Gemini, DeepSeek, Mistral, and anything reachable through LiteLLM
- Repository context: AGENTS.md files and Agent Skills feed the review, describe, and improve tools by default
- Configurable prompting: JSON-based prompts let you retune review categories and behaviour per team
- Restricted mode: runs with reduced GitHub permissions, with GitHub Checks available as an output target
- Plain-diff provider: a tokenless diff path and CI artifact context injection arrived in version 0.40.0
