mu: Detailed Introduction

[ikaijua]

mu: Detailed Introduction

A Pi-based agent that delegates routine decisions to a smaller judge and records its verdicts.

• Audience: Users who want visibility into context selection, approvals, and completion decisions.
• Cost: MIT software; main-model and judge usage may incur separate charges.

Table of Contents

Getting started

1. What it is and who it is for
2. Access and setup: your first session

Usage and configuration

3. A concrete task and how to check it
4. Costs: software, models, and execution
5. Configuration and extensions

Limits and references

6. Permissions, data, and limitations
7. Frequently asked questions
8. Official references

Use the contents to jump to each section. If an in-page link does not respond, open it in a new tab.

[ikaijua]

What it is and who it is for

In one sentence: A Pi-based agent that delegates routine decisions to a smaller judge and records its verdicts.

Who it suits: Users who want visibility into context selection, approvals, and completion decisions.

What you can do:

• Work through a terminal or desktop app.
• Inspect context-selection and approval decisions.
• Compare judges in shadow and active modes.

Where to start: Download the desktop app from Releases, or use Node.js 22.19+ for the CLI.

Cost: MIT software; main-model and judge usage may incur separate charges.

[ikaijua]

Access and setup: your first session

Download the desktop app from Releases, or use Node.js 22.19+ for the CLI.

npm i -g mu-agent
mu setup
mu

The desktop wizard connects a model and folder; CLI setup configures the main model and judge.

[ikaijua]

A concrete task and how to check it

Reproduce and fix one failing test while inspecting the Judge tab, /status, or mu ledger.

Example:

Fix this failure without weakening the assertion. Report the exact test command and remaining failures.

Compare actions with the decision log: a completion verdict is not test evidence. Observe in shadow mode before enabling unfamiliar decision points.

[ikaijua]

Costs: software, models, and execution

Check the provider, network requests, and cost of the judge as well as the main model. Temporary free access to a hosted judge is not a permanent entitlement.

Local judges trade external requests for local compute and memory. Hosted fallbacks can still receive some decision inputs.

[ikaijua]

Configuration and extensions

Decision points can be active, shadow (logged without applying the verdict), or off.

{ "modes": { "default": "shadow" } }

Use ~/.mu/agent/mu.json. Individual points can select different judges; CLI and desktop share settings and sessions. Setup guide

[ikaijua]

Permissions, data, and limitations

/permissions changes the permission mode. Default Judge mode is not approval for every action: some project actions run directly, other actions are evaluated by the judge, and uncertainty escalates.

A probabilistic verdict is not deterministic isolation. Main-model context and judge inputs may reach different providers.

[ikaijua]

Frequently asked questions

Must I configure a judge? Not necessarily, but understand the default data destination.

Does shadow block actions? It records judgments without applying them; permission settings are separate.

How do I debug behavior? Use mu doctor, /status, and mu ledger to identify the relevant decision point.

[ikaijua]

Official references

• Repository
• Getting started
• Judges
• Desktop downloads