mini-swe-agent

mini-swe-agent is a radically minimal, MIT-licensed AI coding agent built by the Princeton and Stanford team behind SWE-bench and SWE-agent, implemented in roughly 100 lines of Python.

The agent deliberately strips away the tool-calling interfaces and stateful shell sessions used by most coding agents, relying solely on bash commands executed independently via subprocess calls, which makes it trivial to sandbox in Docker, Podman, Singularity, or Apptainer. Its message history is completely linear, meaning every agent step simply appends to the conversation, giving a clean 1:1 mapping between trajectory and LLM input, which is ideal for debugging or fine-tuning.

Model-agnostic via LiteLLM, OpenRouter, and Portkey, supporting any LLM including local models for privacy-sensitive use.

Ships with both a simple CLI (mini) and a visual TUI (mini -v) with a built-in trajectory browser.

Supports batch inference and Python bindings for programmatic use.