kitio.aiAdded Nov 5, 2026
mini-SWE-agent
FreeMinimal open-source coding agent that works through bash alone
What is mini-SWE-agent?
mini-SWE-agent is a deliberately small coding agent from the team behind SWE-bench and SWE-agent. Its core is about 100 lines of Python, it only uses bash commands, and it reports scores above 74% on SWE-bench Verified.
The agent skips tool-calling interfaces and keeps a single linear message history, so the trajectory you inspect is exactly what the model saw. Each action runs as an independent subprocess, which makes it simple to swap in Docker, Podman, Singularity, Bubblewrap or other sandboxes and to scale runs out.
Models are reached through LiteLLM, OpenRouter, Portkey and similar routers, so almost any hosted or local model works. Researchers use it as a baseline for benchmarks like SWE-bench and ProgramBench, while developers can run the mini command interactively or import it as a Python library.
Features
- Bash-only agent loop
- Works with any model via LiteLLM
- Docker and sandbox environments
- SWE-bench batch runner
- Trajectory inspector
- Python bindings
Tags
Releases
+ Submit new release- Listed
mini-SWE-agent was added to the directory.
Organization
SWE-agent team (Princeton and Stanford)
🇺🇸 United States
Listed by kitio.ai · unverified owner
Pricing
- Pricing model
- Free
- Paid options from
- 100% free
Sponsored
Your AI tool here
A Featured launch puts your tool in this spot and on the home page for a week.
Top mini-SWE-agent alternatives
View allCoding
Free open-source editor with built-in AI agents for coding
Compare mini-SWE-agent
If you liked mini-SWE-agent
OpenAI's coding agent in the cloud, CLI, IDE and desktop app
Freemium · $8/moCoding50 views41 clicks to the websiteMade in United StatesAnthropic's agentic coding tool for terminal, IDE and web
From $20/moCoding153 views115 clicks to the websiteMade in United StatesOpen-source, model-agnostic coding agent for the terminal
Freemium · $10/moCoding5 views4 clicks to the websiteMade in United StatesOpen standard for connecting AI apps to tools, data and workflows
FreeCoding1 views1 clicks to the websiteAWS coding assistant for the IDE and CLI, formerly CodeWhisperer
Freemium · $19/moCodingMade in United States
Pros and cons
Pros
- Small codebase that is easy to read and modify
- Model-agnostic through common LLM routers
- Strong SWE-bench results for its size
Cons
- No graphical interface
- You pay for your own model API usage
- Aimed more at research than everyday editing
Prompts & results
No prompts yet. Add your own prompt and output to help others use this tool.
Share a prompt
Reviews
Q&A
What is mini-SWE-agent?
mini-SWE-agent is a deliberately small coding agent from the team behind SWE-bench and SWE-agent. Its core is about 100 lines of Python, it only uses bash commands, and it reports scores above 74% on SWE-bench Verified.
Can mini-SWE-agent run with local models?
Yes. The documentation has a section on local models, and any model reachable through LiteLLM or similar routers can be plugged in.
How does mini-SWE-agent sandbox its commands?
Every action is a separate subprocess call, so it can be pointed at Docker, Podman, Singularity, Bubblewrap and other supported environments instead of your local shell.
How much does mini-SWE-agent cost?
mini-SWE-agent is 100% free.
Is there a free version of mini-SWE-agent?
Yes. mini-SWE-agent is completely free to use.
What are the main features of mini-SWE-agent?
Bash-only agent loop, Works with any model via LiteLLM, Docker and sandbox environments, SWE-bench batch runner, Trajectory inspector, Python bindings.
Who makes mini-SWE-agent?
mini-SWE-agent is made by SWE-agent team (Princeton and Stanford).
What are the best alternatives to mini-SWE-agent?
Popular alternatives include Marmot, Stakpak, Butterfish.