mini-swe-agent
Princeton NLPmini-swe-agent is a minimal software-engineering agent whose agent class is about 100 lines of Python. It has no tools other than bash, runs every action as an independent subprocess and keeps a linear message history, which makes it simple to run in Docker, Podman, Singularity and other sandboxes and to use as a baseline for benchmarks and model training. Any model works through LiteLLM, OpenRouter or Portkey. The Princeton and Stanford team behind SWE-bench and SWE-agent builds it.
Openness
5 high confidence- license
- MIT (OSI)
- source
- public
- core features withheld
- no — the tree at 04d809c holds 221 paths with one MIT LICENSE.md and no ee/, enterprise/, commercial/ or proprietary/ path; it is an academic project from the Princeton and Stanford team behind SWE-bench and SWE-agent, and the README describes no hosted service or paid edition
mini-swe-agent is MIT licensed, and the whole agent installs from its public repository or PyPI. It is a research project with no hosted service or paid edition, so nothing is held back.
- https://pypi.org/pypi/mini-swe-agent/json recorded 2026-11-12
PyPI metadata for mini-swe-agent 2.4.6, uploaded 2026-07-23, carries the MIT License classifier and names https://github.com/SWE-agent/mini-swe-agent as its repository.
- https://raw.githubusercontent.com/SWE-agent/mini-swe-agent/main/LICENSE.md recorded 2026-11-12
The LICENSE.md is the MIT License, "Copyright (c) 2025 Kilian A. Lieret and Carlos E. Jimenez".
- https://raw.githubusercontent.com/SWE-agent/mini-swe-agent/main/README.md recorded 2026-11-12
The README says it is "Built by the Princeton & Stanford team behind SWE-bench, SWE-agent, and more" and wants "to be a hackable tool, not a black box"; it describes no hosted or paid tier.
Adoption
5 medium confidenceAdoption is read from PyPI downloads of mini-swe-agent over the trailing month. Automated evaluation and CI runs add to the count, so it measures installs rather than people.
- https://pypistats.org/api/packages/mini-swe-agent/recent recorded 2026-11-12
last_month 11,605,679 downloads for mini-swe-agent (last_week 3,349,901, last_day 577,948).
Capability
3 medium confidencemini-swe-agent reduces a coding agent to bash and a linear history, and its README reports SWE-bench Verified results that depend on the model it drives. It sits level with SWE-agent, the same team's earlier and more configurable agent, whose own README now recommends mini-swe-agent instead.
- https://raw.githubusercontent.com/SWE-agent/mini-swe-agent/main/README.md recorded 2026-11-12
The README says the agent "Does not have any tools other than bash", executes actions with subprocess.run, has "a completely linear history", supports local, docker/podman, singularity/apptainer and bublewrap environments, and "Scores >74% on the SWE-bench verified benchmark".
Verified 2026-11-12