SWE-agent

SWE-agent is a command-line coding agent for developers and researchers attempting fixes in GitHub repositories. It runs the agent on your machine and supports local Ollama models through a configured endpoint, or cloud APIs such as OpenAI and Anthropic. Cloud-backed runs send model requests to the selected provider.

Issue repair is its main use, but it also supports custom tasks, competitive coding challenges and cybersecurity vulnerability research. The model chooses how to use the available tools rather than following a fixed sequence for every issue.

The Python codebase is designed for researchers to modify agent behavior and evaluate it on SWE-bench. YAML configuration controls the agent, and the documentation covers command-line tasks and benchmarking. This makes it relevant to experiments with coding agents as well as attempts to repair individual repository issues.

SWE-agent is open source under the MIT license. Researchers at Princeton University and Stanford University maintain the project.

Local models require matching tool-call parsing and history settings. The selected model’s capabilities affect which configuration works.