SWE-agent
An open-source AI coding agent that uses GPT-4o or Claude Sonnet 4 to tackle GitHub issues, custom tasks and security challenges. MIT licensed.
SWE-agent is an open-source AI coding agent that lets a language model use tools to attempt fixes for issues in GitHub repositories. It's aimed at developers and researchers studying how agents handle real software tasks. The project is in maintenance-only mode: mini-swe-agent has superseded it, and the maintainers recommend that successor for new users.
You choose the language model, with GPT-4o and Claude Sonnet 4 among the named examples. The agent gives that model broad freedom to decide how to use its tools, rather than restricting it to a fixed sequence of actions. Its scope extends beyond issue fixes to custom tasks and finding cybersecurity vulnerabilities. With a vision-capable model, it can also process images attached to GitHub issues.
For researchers, the Python codebase is designed to be easy to modify, with documented configuration and support for custom tools. SWE-bench benchmarking is part of the project's documented workflow, so it suits experiments that compare agents on repository repair tasks. The MIT license permits modification and reuse.
SWE-agent also has an EnIGMA mode for offensive cybersecurity capture-the-flag challenges, and it can tackle competitive coding problems. Researchers at Princeton University and Stanford University built the project; their main development effort is focused on mini-swe-agent.
