mini-swe-agent is a radically simple AI software engineering agent built by the Princeton & Stanford team behind SWE-bench and SWE-agent. It uses only bash — no custom tools, no giant monorepo — yet scores >74% on SWE-bench verified, making it one of the most capable and accessible coding agents available.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/swe-agent/mini-swe-agent/llms.txt
Use this file to discover all available pages before exploring further.
Quickstart
Install mini-swe-agent and run your first task in minutes
CLI reference
Learn all commands and modes of the
mini interactive CLIConfiguration
Customize agent behavior with YAML config files
Python bindings
Embed mini-swe-agent in your own Python code
Why mini-swe-agent?
Most coding agents are either overfitted research artifacts or UI-heavy frontend products. mini-swe-agent takes a different approach: it is small enough to understand at a glance, yet powerful enough for real-world use.Minimal
~100 lines for the core agent class. No tool-calling interface — just bash.
Performant
74% on SWE-bench verified with Gemini 2.5 Pro. Starts faster than Claude Code.
Flexible
Runs locally, in Docker, Singularity, Bubblewrap, Modal, and more.
Compatible
Supports every model via litellm, OpenRouter, Portkey, and Requesty.
Hackable
Subclass agents, environments, and models to build your own variant.
Debuggable
Linear message history means trajectories are exactly what the LM sees.
Get started
See the quickstart guide for all installation options, includinguvx and pipx for isolated environments.
Explore
Environments
Run in Docker, Singularity, Bubblewrap, Modal, and ConTree
SWE-bench evaluation
Batch-evaluate on SWE-bench verified and other benchmarks
Extend mini
Build custom agents, environments, and models
FAQ
Answers to common questions about mini-swe-agent