mini-SWE-agent

mini-SWE-agent is a lightweight AI coding agent built in approximately 100 lines of Python code. It is designed as a minimalist tool that relies exclusively on a bash interface rather than custom tool-calling mechanisms. The software executes actions via independent subprocesses with a linear message history to provide a simplified developer experience.

• Relies exclusively on a bash interface rather than custom tool-calling mechanisms
• Executes actions via independent subprocesses using subprocess.run
• Maintains a completely linear message history where every step appends to messages
• Supports local environments, Docker, Podman, Singularity, Apptainer, Bubblewrap, and Contree
• Compatible with all language models via Litellm, OpenRouter, Portkey, and more
• Provides a command line interface, batch inference, and a trajectory browser
• Offers Python bindings for programmatic usage

Pros

• Provides an extremely simple control flow and codebase that is easy to understand at a glance
• Offers broad model compatibility by avoiding specialized language model tool-calling interfaces
• Simplifies debugging and fine-tuning through a completely linear message history
• Enables stable sandboxing and scaling through independent subprocess action execution

Cons

• Lacks specialized tools outside of bash, relying entirely on the language model to figure out tasks
• Does not maintain a stateful shell session running between actions
• Requires checking documentation or migration guides when moving between software versions

This tool fits developers, researchers, and technical users looking for a hackable command line coding agent with a minimal codebase and transparent control flow. It is well-suited for those doing fine-tuning or reinforcement learning who want to avoid overfitting to complex agent scaffolds. It is a poor fit for users who want agents with pre-built specialized custom tool interfaces or stateful interactive shell sessions.

Unlike heavier or more complex AI coding agents that implement custom tool-calling mechanisms and dedicated interfaces for every action, mini-SWE-agent strips the design down to a bash-only approach. It replaces stateful shell sessions with independent subprocess executions and linear message histories. This radically simpler structure puts the language model directly in the center of attention rather than relying on heavy agent scaffolds.

Litellm, OpenRouter, Portkey, Local environments, Docker, Podman, Singularity, Apptainer, Bubblewrap, Contree

mini-SWE-agent is worth trying as a default choice for developers and researchers who want a fast, transparent, and hackable coding agent. Its minimal codebase and bash-only design make it exceptionally useful for debugging, benchmarking, and custom model integration. Users who require specialized pre-built tool interfaces or a graphical user interface frontend should look elsewhere or use the full swe-agent.