Mi - 30-line zero-config CLI agent for bug fixes + refactoring
by•
mi is an autonomous coding agent that fits in a single JavaScript file. no framework, no dependencies beyond Node builtins, works with any OpenAI-compatible API — OpenAI, Ollama, local models, whatever you have.
the core is a loop: call the llm, check if it wants to use tools, execute them, feed results back, repeat. two built-in tools — bash (full system access) and skills (markdown playbooks loaded on demand) — are enough for it to read repos, write code, run tests, and debug failures.
Replies
Facts
mi was born as a curiosity piece.
how small can an agentic harness be and remain useful? turns out, very small. Claude Code is ~500k lines of code, OpenCode is ~100k, mi is 30 lines, yet, it has all the features expected from agentic harness at this stage and then some.
mi is a loop, two tools, and an LLM that you plug in.
mi is optimised for running with local LLMs, it's concise and to the point. it can extend itself to solve task with fewer tokens. this launch on Product Hunt is paired with launch of --sandbox mode, mi will run itself in a container, away from your host and harming you with downstream issues.
thank you!