mu

mu (μ): most of what a coding agent decides isn't code. mu hands 40 kinds of yes/no calls to a small model: fewer wasted tokens, fewer approvals on autopilot, f

What is mu?

mu (μ): most of what a coding agent decides isn't code. mu hands 40 kinds of yes/no calls to a small model: fewer wasted tokens, fewer approvals on autopilot, fewer runs that stop halfway. Built on pi.

Overview of mu

What it is

mu (μ) is a tool that handles yes/no decisions in coding agent loops by delegating them to a small, fast model. It is built on pi. In Terminal-Bench 2.1, with the same model and tasks, mu passed 68.2% of attempts while pi passed 58.7% (+9.5 points, 95% interval +3.8 to +15.2). mu aims to reduce wasted tokens, commands approved on autopilot, and runs that stop halfway.

Key features

  • 40 decision points in the agent loop, 34 of which act on their verdicts out of the box.
  • Default judge model is Jev, hosted by TypeSafe; alternatives include mude 1.0 (local), Hugging Face models, or any LLM.
  • Test-log trimming (off by default) folds exact repeats: cuts 40% of characters in Jest failure logs and 19.5% across 71 real failing test logs.
  • Shadow mode records what the judge would do without acting on it, for comparison.
  • Every verdict is logged in a local ledger (mu ledger, /status, or the desktop app's Judge tab).
  • Available as a command-line tool (mu) and a desktop app for macOS, Windows, and Linux.

Use cases

  • Reducing token waste by having the judge read long output chunk by chunk and archive irrelevant content behind a pointer.
  • Preventing autopilot approvals: in Judge mode (default), the judge lets through commands the task clearly needs and asks about the rest.
  • Catching runs that stop short or unverified completions, sending the agent back to finish.
  • Screening web pages and MCP output for instructions aimed at the AI before the model sees them.
  • Comparing different judges before relying on one, using shadow mode.

Getting started

Desktop app: download the installer for macOS, Windows, or Linux from Releases, open it, and sign in with ChatGPT or paste an API key. Command line: bash npm i -g mu-agent # Node.js 22.19 or newer mu setup # connect a model mu # start a session in the current directory You do not need a judge key to start; the free Jev on OpenCode Zen answers until you set one.

License and pricing

The project is open source under the MIT license.

Based on the project's public README and repository metadata.

Add this badge to your README

Featured on YunoDB Badge styles & help

Maintain this project? Show that it is featured on YunoDB.

Markdown[![Featured on YunoDB](https://yunodb.com/badge/qybaihemu.svg)](https://yunodb.com/tools/qybaihemu?utm_source=github&utm_medium=badge)
HTML<a href="https://yunodb.com/tools/qybaihemu?utm_source=github&utm_medium=badge"><img src="https://yunodb.com/badge/qybaihemu.svg" alt="Featured on YunoDB" height="20"></a>