This project provides a minimalist, 100-line AI agent designed for researchers and developers to benchmark, fine-tune, or deploy AI coding assistants with minimal overhead. It achieves 65% accuracy on the SWE-bench benchmark using Claude Sonnet 4, offering a simple, hackable alternative to more complex agent frameworks.

How It Works

mini-SWE-agent leverages a radically simple design, eschewing complex tools and stateful shell sessions. It relies solely on bash commands executed via subprocess.run, allowing any language model to interact with the environment. This approach simplifies sandboxing, debugging, and fine-tuning by maintaining a linear execution history and ensuring each action is independent.

Quick Start & Requirements

• Install: pip install uv && uvx mini-swe-agent or pip install pipx && pipx ensurepath && pipx run mini-swe-agent
• Prerequisites: Python. No specific model or API keys are mandated by the agent itself, but a compatible LLM is required for operation.
• Docs: Quick start

Highlighted Details

• Scores 65% on the verified SWE-bench benchmark.
• Minimalist design: ~100 lines of Python code.
• No tools beyond bash; relies on LLM's ability to use the shell.
• Supports local execution and containerized environments (Docker, Podman).

Maintenance & Community

Developed by the team behind SWE-bench and SWE-agent from Princeton and Stanford. Further details on community and contribution can be found in the project's documentation.

Licensing & Compatibility

The repository does not explicitly state a license in the provided README. Users should verify licensing for commercial or closed-source use.

Limitations & Caveats

The agent's functionality is limited to what can be achieved through bash commands, meaning complex operations requiring specific tools or APIs must be handled by the LLM's ability to generate appropriate shell commands.