mini-swe-agent
mini-swe-agent is a deliberately minimal AI software-engineering agent from the Princeton and Stanford team behind SWE-bench and SWE-agent. The core agent class is around 100 lines of Python, and the agent has no tools other than bash: it does not even use the model's tool-calling interface, so it works with essentially any model.
It suits researchers who need a simple, transparent baseline for coding-agent experiments and benchmarks, and developers who want a lightweight agent that solves GitHub issues or helps in the terminal. Despite its size, the project reports scores above 74% on SWE-bench Verified, and it is used by organisations including Meta, NVIDIA, IBM, Princeton and Stanford. It runs in local environments as well as Docker or Podman, Singularity/Apptainer and other sandboxes, and supports models through LiteLLM, OpenRouter, Portkey and more. Version 2 is current; a migration guide covers changes from v1.
Getting started: run it without installing via `uvx mini-swe-agent` or `pipx run mini-swe-agent`, or install with `pip install mini-swe-agent`, then configure a model key.
Features
Why choose it
Trade-offs
Maintenance check
- Last commit
- 2026-09-03
- Commits, last 90 days
- 3
- Releases, last 90 days
- 0
- Latest release
- v2.4.6 (2026-07-23)
- Open issues
- 125
- Recent issues with at least one comment
- 5/12
Oldest week on the left; the latest four weeks are orange. Source: GitHub commit activity.
Status rule: archived on GitHub = Archived; last commit within 90 days = Active; 91 to 365 days = Slowing; older = Stale. All figures come from the GitHub API on the date shown.
Most discussed open issues
- #797 Why use `||` between `stop` and `rm` when cleanup a docker environment?7 comments · 2026-03-26
- #1003 [Bug] litellm 1.106.0 breaks every tool-using run on base installs: ModuleNotFoundError: No module named 'mcp'4 comments · 2026-10-18
- #953 RFC: deterministic execution boundaries for tool calls (task text is untrusted input)4 comments · 2026-08-29
- #829 Bash tool launches Dash on Ubuntu environments4 comments · 2026-05-18
- #546 plans to integrate remote runtime execution using swe-rex?3 comments · 2025-09-27
Issue titles are quoted from GitHub. Open the link to check whether a problem still applies to the version you use.
Who it fits
- You are researching autonomous software engineering agents and want a radically minimalist, 100-line Python codebase that scores >74% on SWE-bench.
- You want to build or hack custom automated issue-resolution loops without the bloat of massive monorepos or complex tool-calling abstractions.
- You need a polished, interactive daily driver with full terminal UI, undo history, and syntax diff overlays (Aider or Claude Code are built for daily coding).
- You require multi-agent orchestration or enterprise-scale semantic codebase knowledge graphs.
How it differs
vs learn-claude-code★ 79.4kmini-swe-agent is a hyper-compact benchmark-tested baseline for autonomous bug fixing; learn-claude-code is a progressive step-by-step tutorial for building modern coding agents.
vs ruflo★ 79.8kmini-swe-agent pursues extreme minimalism in ~100 lines of Python; ruflo is an enterprise-scale swarm platform with 15 agent roles and vector memory.
Compatibility
Quick start
Use cases
Alternatives
Comments
No comments yet. Be the first!