What it does

mu is a coding agent built on pi that adds what it calls a judgment kernel. Many of a coding agent's decisions concern running the session rather than writing code, such as what the context holds, whether a command is safe and when to stop. mu hands these to a small, fast judge model that is asked one narrow question at a time, so the main model can spend its attention on the work. The judge is consulted at dozens of decision points in every turn. Examples include which chunks of a long tool output enter the context, which saved lessons apply to a task, whether a rule-flagged command was actually requested, and whether web or MCP output carries instructions aimed at the AI. Each decision point can be active, shadow (asked and logged without acting) or off, and every verdict is recorded in a ledger. The default judge is Jev; a local 322M-parameter judge called Laya, a classifier or a general LLM can be assigned to a decision point instead. mu runs as a command-line tool installed with npm (Node.js 22.19 or newer) and as a desktop app for macOS, Windows and Linux. The desktop app adds panels for judgments, a multi-agent hive, files, preview and a built-in browser. The page describes the project as in early development, with 0.1.x pre-releases whose interfaces and formats are not yet stable.

Capabilities

Small judge model makes routine session decisions so the main model can focus on the work
Long tool output is admitted in pieces, and outdated results are cleared from the context
Web page and MCP output that carries instructions aimed at the AI is withheld
Every verdict is logged in a ledger, and decision points can run in shadow mode
Hive of two to six read-only bees that share findings on a common board
Command line plus a desktop app for macOS, Windows and Linux

When to use it

Long coding sessions where tool output keeps filling the context
Comparing judge models in shadow mode before letting one act
Agent work that reads web pages or MCP servers whose output should be screened
Continuing existing Claude Code or Codex CLI conversations in another agent