qybaihe/mu

What is it?

Key features

• 01Routes permission checks, context admission and turn-completion decisions to a judge model.
• 02Supports Jev, the local mude 1.0 model, Hugging Face decision models, classifiers and any LLM as judges.
• 03Lets judges be chained in tiers, cheapest first, with a fallback when none answers.
• 04Logs every verdict with probabilities and latency in a local ledger.
• 05Offers a shadow mode to compare judges without changing behavior.
• 06Provides a desktop app and a command-line tool that share sessions and settings.

Use cases

• Running a coding agent that asks for fewer confirmations while still catching risky commands.
• Keeping long test logs and tool outputs out of the model's context to save tokens.
• Comparing different judge models on your own sessions before choosing one.
• Running routine coding-agent decisions locally on a Mac with Apple silicon.

Quick start

1. 1

   Install the CLI from npm (Node.js 22.19 or newer)

   npm i -g mu-agent
2. 2

   Connect a model

   mu setup
3. 3

   Start a session in the current directory

   mu

Things to know

• !Only the first step is a desktop installer; the CLI needs Node.js 22.19 or newer.
• !The default judge is a hosted model, and mu's judge decisions depend on it unless you configure another judge.
• !Several of the 40 decision points are off or in shadow mode by default.

In collections

Topics

FAQ

What is mu?

A coding agent built on pi that routes routine decisions, like permission checks and context pruning, to a small judge model so the main model can focus on code. mu is a coding agent that sends the many non-code decisions in a session to a small, fast judge model. Each decision point asks one bounded question, such as whether a command is safe, whether a tool output is worth keeping in context, or whether the work is actually finished. The main model keeps its attention on the task, and every verdict is recorded in a local ledger with its probabilities and timing.

Is mu free to use? What license does it have?

mu is open source under the MIT license. Check the license terms before using it in commercial products.

Is mu still maintained?

Actively maintained. The last commit to mu was on Oct 24, 2026.

What can I use mu for?

Running a coding agent that asks for fewer confirmations while still catching risky commands. Keeping long test logs and tool outputs out of the model's context to save tokens. Comparing different judge models on your own sessions before choosing one. Running routine coding-agent decisions locally on a Mac with Apple silicon.

How do I install mu?

Install the CLI from npm (Node.js 22.19 or newer): npm i -g mu-agent Connect a model: mu setup Start a session in the current directory: mu

Where can I get mu?

The source code and installation instructions are on GitHub: https://github.com/qybaihe/mu