miniagent

miniagent is a terminal coding agent designed for local models. It runs shell commands, edits files, searches and browses the web, and maintains work across long tasks, all driven by a model running on the user's own machine without an API key and with nothing leaving the computer. It was built and tuned against Spark-X2.5-4B in LM Studio on Windows 11, and it speaks the ordinary OpenAI-compatible API, so llama.cpp, Ollama, vLLM, SGLang or a remote endpoint work just as well. It also drives Qwen3.8-27B on a free Kaggle TPU, served by kaggle-tpu-lab/ in the repository. The agent runs on the user's PC using only the standard library, and includes benchmark tasks and checkers, settings for a local SearXNG for web search, and the model side on a free Kaggle TPU. Requirements include Windows 11 (Windows 10 works too), Python 3.8+ with no third-party packages, a local model server such as LM Studio, and optionally Docker for a local SearXNG. The agent can be pointed to different endpoints with flags or environment variables, and it supports profiles like Ternary Bonsai 2 27B. Web search is optional but recommended, using a local SearXNG to avoid rate limits. The Kaggle TPU integration allows running Qwen3.8-27B with vLLM on Kaggle's free TPU v5e-8, with the agent running locally and only the model on Kaggle.