What does a Jev-backed turn cost me in latency and tokens?

Measured by the authors on their own sessions, one warm question took about 0.3s over HTTP/2, and 16 chunks of tool output were judged in a single request in 0.44s with the state billed once. Verdicts, probabilities, and timings go to the ledger.

Can I keep everything on my machine, and what do I lose?

Laya is a 322M-parameter local judge that never touches the network and downloads nothing without consent. The README calls it reliable on simple predicates but weaker on meta-judgments, and suggests running it in shadow beside Jev before giving it a real decision point.

What does shadow mode on a decision point actually do?

A shadow decision point is asked and logged but changes nothing. That is the documented way to compare judges on your own traffic before switching one on, and the ledger is where you read the comparison.

Can the judge wave through a dangerous command?

No. Rules catch a dangerous-looking command before any judge is consulted, and the judge only vouches that you asked for it. In Jev approves mode only what the judge is sure of runs, and everything else asks you.

How do I install and start it?

The desktop app ships installers for macOS, Windows and Linux; the command line installs with npm i -g mu-agent on Node.js 22.19+, then mu setup connects a model and mu starts a session.

Do I need a TypeSafe key to use the judge?

No. Until you set a judge key, the free Jev on OpenCode Zen answers, and every decision point acts on its verdicts from the start. You can also record only what a judge would do.

Can a decision point use a judge other than Jev?

Yes. Each point can name jev, a local laya, classifier: / , llm: / , or a cascade such as laya,jev, and each can be set active, shadow or off.

CommunityJev Agents & MCP

Mu

A coding agent built on pi that hands its non-code decisions — what enters the context, whether a command is safe, whether the work is done — to a small judge model at 35 decision points per turn.

See highlights

Open github.com

Stars: 488
Language: TypeScript
License: MIT
Published by: Community
Author: qybaihe
Last verified: Oct 2026

qybaihe/mu

User repository on GitHub

View on GitHub

mu (μ): a coding agent that thinks before it acts. A small, fast judge makes the routine calls, the big model does the work. Built on pi and AionUi.

Forks: 50
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

mu μ · Only what's needed. A coding agent with a judgment kernel. Built on pi . English · 简体中文 · 繁體中文 · 日本語 · 한국어.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest

Works forJev for Classification Jev for Tool calling Jev for Routing

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Every decision point is active, shadow, or off, and can name its own judge: jev, laya, llm:<provider>/<model>, or a cascade.
Rules stay the floor: a command the rules flag is caught first, and the judge only vouches that you asked for it.
The desktop app judges page picks a hosted judge by key and defaults to the first one whose key is set.
Exact repeats in failing test logs are folded without calling a model, and the markers expand byte for byte.

Watch out

The README labels this early development with 0.1.x pre-releases and warns that names, settings, and formats may still change. Latency figures come from the authors' own sessions, and the test-log numbers come from the repository's own replay rather than an outside measurement. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

agentconnect

1.5k

Open-source platform that puts ACP-compatible agents such as Claude Code, Codex, and Pi inside the team's Slack, Telegram, Discord, and GitHub workflows, with reusable Jev decisions for routing, response gating, and reviewer selection.

Jev Agents & MCP#community#jevbest

Community

agentconnect-md/agentconnectTypeScript

duet-agent

Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.