Mini SWE Agent

Mini SWE Agent is a command-line AI coding assistant for developers tackling GitHub issues and everyday coding tasks. Its small Python core also suits researchers who want to inspect and modify an agent for model evaluations.

• Shell-based coding: The agent uses bash to carry out tasks, so it can work with models that don't support a tool-calling API.
• Model choice: It supports local models and connects to model services through LiteLLM, OpenRouter and Portkey.
• Run inspection: A trajectory browser lets you review agent steps, and the run history matches the messages sent to the model, making prompts and decisions easier to examine.
• Research workflows: Batch inference and Python bindings support model comparisons and use from other programs; the simple agent design provides a baseline for fine-tuning and reinforcement learning.

The agent runs on your machine or in container and sandbox environments, including Docker, Podman and Singularity/Apptainer. Each shell action runs independently, so tasks don't share a persistent shell session. The project is open source under the MIT license. Model processing depends on your chosen backend: cloud services receive model requests even when the agent itself runs locally.