μ · Only what's needed.
mu is a coding agent. A small, fast judge called Jev makes dozens of small decisions for the big model in every turn: what stays in the context, whether a command is safe, whether the work is really done. The big model keeps its attention for the work.
$ npm i -g mu-agent
Needs Node.js 22.19 or newer. Then run mu in a project folder; the first time, mu setup connects a model.
v0.1.7 · The macOS app is signed and notarized by Apple; the Windows app is not signed yet, so SmartScreen shows a warning. The app carries mu inside it: no Node.js to install. Windows on Arm, Linux on Arm and every other download are under Releases.
MU · A REAL TURN
You ask for a plan, and the context carries a whole toolbox. Models write better code every month, and their reports get harder to read. mu splits the work: Jev judges, the big model works, and a model picked for plain speaking tells you where things stand.
4 .toLowerCase() 5 .replace(/[^a-z0-9]+/g, "-") -6 .replace(/^-/, ""); +6 .replace(/^-+|-+$/g, ""); 7 } $ cd /Users/Shared/mu/slugify && npm test 2>&1 ... (10 earlier lines, ctrl+o to expand) ℹ fail 0 ℹ cancelled 0 ℹ skipped 0 ℹ todo 0 ℹ duration_ms 84.5005 Took 0.3s All 4 tests pass now. The bug was in src/slugify.js — the final .replace() only stripped a leading dash (/^-/) but not a trailing one. Changed it to /^-+|-+$/g to strip dashes from both ends. ──────────────────────────────────────────────────────────────────────────────────── ──────────────────────────────────────────────────────────────────────────────────── ~/slugify (master) ↑6.2k ↓643 R19k CH94.0% $0.034 (sub) 2.5%/200k (auto) claude-sonnet-4-6 • medium jev-1.13-free active: multi_step_task -> heavy 752ms Permissions: Full access
When you speak
Is this a new task, a follow-up, or a correction? Should a message that arrives mid-run interrupt now? Do skills and tools need to come in, or can one sentence answer it?
This is a real turn. "Run the tests and fix the failing ones" comes in, and in under a second Jev answers: a multi-step task, heavy gear. Skills and MCP servers the task does not need stay out of the context until it does.
When a tool answers
A test run or a search easily returns hundreds of lines. Jev judges chunk by chunk what matters now: the failing cases stay, logs that repeat themselves go. What stays out is archived behind a pointer, ready when it is needed.
In web pages and MCP results, instructions aimed at the model are withheld before it reads them.
Before acting
Before each command runs, Jev judges whether it does something that cannot be undone, and whether it crosses a rule you set. There are three permission modes: Full access, Jev approves and Minimal permissions, switchable at any time. Under Jev approves, you are asked only what needs you; the rest runs.
A verdict changes what the model does next, never whether it asks you.
When the turn ends
When the model says it is done, mu checks whether anything verified that, whether the turn drifted, and whether to go back to a checkpoint.
/goal <condition> keeps the agent working until the condition holds. A big model checks whether it does, with Jev as the fallback; when the same error comes back again and again, Jev judges whether the approach is a dead end.
The hive
When several agents work on one hard problem, handing out the work is easy; talking is hard. Say nothing, and each walks into the same dead end. Say everything, and every context fills with the others' chatter.
In mu's hive, Jev judges every finding: is it worth sharing, and does it matter to this bee? Only then does it reach another bee as context. When a later finding overturns an earlier one, the correction goes to every bee that heard it.
The plain-language board
Models get stronger, and their reports read more and more like notes for a machine. In mu the working model keeps working its own way. Every few tool calls Jev reads the scene, picks what is really news, and hands it to a model chosen for speaking plainly, which tells you where things stand.
When a run ends, you see not just "done" but what was actually done.
WHAT WE LEFT OUT
A big model's attention is limited, and so is its context. These are the things mu chose not to do, or chose to do differently.
A verdict from Jev changes what the model does next, never whether it comes back to ask you.
The first message is judged: small talk, a follow-up, or work to do. Skills and MCP servers that are not needed stay out of the context.
When the context grows, stale results are dropped instead of summarized; before the cache expires, mu judges whether a refresh is worth it.
38 decision points, each with its own switch and its own judge: the free Jev, the local Laya, or any LLM.
Without a judge key, the free Jev on OpenCode Zen answers, and mu says so once a day.
The command line is built on pi, with its models, extensions, skills and session tree; the desktop app is built on AionUi.
ALSO
/status, mu ledger, and the app's judgments tab.GET INVOLVED
Report a bug or ask for a feature; the forms ask for what is needed.
Questions, ideas and what you built with mu.
Fix a bug, propose a decision point, measure a judge or translate; good first issues are labelled.
The Jev key giveaway has ended and the claim page is gone. mu now works without a key; to use your own TypeSafe key or another judge, see Judges.
A star helps more people find it. The code, the discussions and every release live in the repository.