AutoCodeRover

Detailed overview

AutoCodeRover is an open-source, program-structure-aware autonomous agent for resolving GitHub issues such as bug fixes and feature additions. It works in two stages: context retrieval, where an LLM uses AST-based code search APIs to locate relevant methods and classes rather than relying on plain string matching, and patch generation, where it drafts a fix from that context. When test suites exist, it applies statistical fault localization to improve repair accuracy. It supports OpenAI, Anthropic, Gemini, Groq, and AWS Bedrock-hosted models, and can run via Docker or a local conda environment against live GitHub issues, local issue descriptions, or SWE-bench task instances. On SWE-bench Lite and Verified benchmarks it has reported pass@1 rates of 37.3% and 46.2% respectively, with each task costing under $0.7. It is aimed at researchers and developers experimenting with autonomous program repair rather than production-ready automated maintenance.

Pricing details

Free and open source; model and infrastructure costs apply.