SWE-agent

Key features

• Point the command line tool at a GitHub issue and the agent attempts a fix on its own, working through the repository until it produces a change.[1]

• Model access goes through LiteLLM, so operators can use GPT-4o, Claude, or other supported providers with their own API keys rather than being locked to one vendor.[1][2]

• The agent works through an agent-computer interface of purpose-built commands for viewing, editing, searching, and executing code, leaving the problem-solving strategy to the language model.[1]

• Every run executes inside a Docker sandbox, and trajectories are recorded so runs can be inspected and replayed afterward.[1]

• Behavior is governed by a single YAML configuration file with documented templates for models, tools, environments, and demonstrations.[1]

• A batch mode runs the agent across SWE-bench task sets for benchmarking, with tooling for competitive runs and output comparison.[1]

Who will love it

• Developers who triage real GitHub issues and want an autonomous first attempt at a fix will get the most from the issue-to-patch workflow.[1][2]

• Agent researchers benefit from the hackable design, recorded trajectories, and batch evaluation support built for SWE-bench experiments.[1][2]

• Python shops comfortable with pip installs, Docker, and bringing their own model keys can self-host the whole setup without paying for software.[1][2]

Who should avoid it

• Setup is moderately involved: Python 3.11 or newer, an editable pip install, Docker for the sandbox, and a configured model API key are all required before the first run.[1][2]

• The agent cannot work offline. Documented runs depend on a hosted language-model API and container images, so network access and provider billing apply.[1][2]

• Maintainers now steer new users toward mini-swe-agent as the simpler successor, so teams adopting SWE-agent itself should expect the ecosystem's momentum to sit elsewhere.[1]

Alternatives and compatibility