mini-SWE-agent

mini-SWE-agent (mini) is a minimal coding agent from the Princeton and Stanford team behind SWE-bench and SWE-agent. The agent class is roughly 100 lines of Python. It has no tools other than bash, keeps a completely linear message history, and runs each action as an independent subprocess.run call rather than in a stateful shell session.

Because it does not depend on the tool-calling interface of language models, it can be used with a wide range of models via litellm, OpenRouter, Portkey and others. It supports local execution as well as Docker/Podman, Singularity/Apptainer, bubblewrap and other sandboxes. The README reports a score above 74% on SWE-bench Verified. It is used as the fixed harness for the SWE-bench "bash only" leaderboard, and the README says it is also used for Ramp SWE-Bench and the DeepSWE evaluation.

It is meant both as a research baseline that keeps attention on the language model rather than the scaffold, and as a hackable command-line tool for everyday use. It is installed from PyPI and run locally. The current major version is v2, which has a migration guide from v1. As a deliberately simple harness, models may perform better in their vendor's own agents, as one YouTube reviewer noted about Claude Code.