mu
Pi-based coding agent where a small judge (Jev by default) answers yes/no, choice and score questions at 38 decision points per turn.
Overview
mu is a coding agent built on pi, available as a CLI and a desktop app built on AionUi. Its premise is that an agent makes hundreds of non-coding decisions per session, such as what stays in context, whether a command is safe, whether fetched content carries injected instructions, and when work is done. mu hands these to a judge, a small fast model answering one bounded question at a time at 38 decision points per turn, so the main model can focus on the work.
Jev is the default judge, returning a probability per answer and logging every verdict in a ledger; a local Laya model or any LLM can take a decision point instead. Examples include tool.risk and tool.approval before commands run, tool.injection to withhold instructions in web pages or MCP output, and tool.admission to decide chunk by chunk what enters context. Rules catch dangerous commands first, and each point can run in shadow mode. It is in early development with 0.1.x pre-releases.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this fetched page or MCP result contain instructions aimed at the model? | yes · no |
| Choice | Should this tool-output chunk enter the context or be archived behind a pointer?(inferred) | admit · archive |
| Yes / no | Was this rule-flagged command actually asked for? | yes · no |
| Yes / no | Was the claimed completion verified? | yes · no |
How it works
- User input passes preflight and task framing decision points.
- Model proposes tool calls; rules and the judge assess risk and approval.
- Tool output is checked for injection and admitted to context chunk by chunk.
- Verdicts are acted on (or recorded only in shadow mode) and logged to the ledger.
Strengths
- Offloads routine calls from the main model, saving tokens and latency.
- Every verdict is logged and inspectable (mu ledger, /status).
- Shadow mode lets you measure the judge before letting it act.
- Free Jev via OpenCode Zen works until you add your own key.
Caveats
- Early development; names, settings and formats may change.
- Judge errors can drop useful context or wrongly block commands.
Getting started
npm i -g mu-agent (Node 22.19+), then mu setup and mu in your project directory.At a glance
- Type
- CLI
- Status
- prototype
- Language
- TypeScript
- Runtime
- CLI, desktop app
- Has code
- yes
- License
- MIT
- Stars
- 504
Use cases and integrations
How curated lists describe it
Coding agents: a pi-based coding agent and desktop app that asks Jev `Noul`, `Choice` and `Score` questions at 38 decision points inside its loop, such as which chunks of a long tool output enter the context, whether a rule-flagged command was actually asked for, whether a fetched page or MCP result carries instructions aimed at the model, and whether a "done" was verified, with each point set to active, shadow or off and every verdict written to a local ledger; in the repository's replay benchmark, goal-aware test-log selection by Jev cut 40-46% of the log without losing a required line.
yibie/awesome-jevA coding agent built on pi and AionUi whose routine calls at 38 decision points are made by Jev.
logicrw/awesome-jev-projectsmu (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.
wh000wh000/awesome-jev-livePi-based coding agent and desktop app where Jev answers `Noul`, `Choice` and `Score` questions at 38 decision points, such as which chunks of a tool output enter the context, whether a rule-flagged command was asked for, whether fetched pages or MCP output carry injected instructions, and whether "done" was verified; code acts on each answer, rules catch dangerous commands first, and each point can run active, shadow or off with every verdict in a local ledger.
cobanov/awesome-jevMu (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.
hellogumbo/awesome-jevCoding agent built on pi with a judgement kernel deciding the routine calls. Five languages.
MrJev/awesome-jev
Frequently asked questions
What does the judge decide in mu?
Routine calls at 38 points per turn, such as command risk, prompt-injection checks, context admission and whether work is really done.
Do I need a TypeSafe key?
Not to start; a free Jev on OpenCode Zen answers until you set your own judge key, and Laya or any LLM can be used instead.
Can I watch without letting it act?
Yes, set the default mode to shadow in ~/.mu/agent/mu.json to only record verdicts.
Related projects
OMP extension keeping one endless chat as a summary tree, with five Jev judge gates for rules, prefetch, compression, tools and recall.
PluginprototypeTypeScriptClaude Code mod that hands small decisions (routing, skill choice, screening, gating, trimming) to Jev or another decision backend.
ProjectbetaTypeScriptAgent context compactor where Jev scores transcript segments to keep verbatim or move behind a recoverable pointer.
ServiceprototypePythonSecurity hook giving Codex, Copilot, Gemini, Cursor, pi, OpenCode and ACP an auto-mode-like guard via three Jev questions.
LibrarybetaTypeScriptJev Sentinel: guard for Pi, Claude Code and Codex that checks tool calls, tool outputs and replies, failing closed.
PluginbetaTypeScriptPre-execution hook for coding agents that lets shell commands run, asks a human, or blocks them, with optional Jev classification.
ProjectbetaJavaScript