by qybaihe
mu (μ): most of what a coding agent decides isn't code. mu hands 40 kinds of yes/no calls to a small model: fewer wasted tokens, fewer approvals on autopilot, fewer runs that stop halfway. Built on pi.
Requires a passing catalog security scan.
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM lists this project for discovery and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
# Add to your Claude Code skills
git clone https://github.com/qybaihe/muGuides for using ai agents skills like mu.
Last scanned: 10/6/2026
{
"issues": [
{
"type": "npm-audit",
"message": "@earendil-works/gondolin: Vulnerability found",
"severity": "high"
},
{
"type": "npm-audit",
"message": "braces: braces vulnerable to stack-exhaustion denial of service through deeply nested patterns",
"severity": "high"
},
{
"type": "npm-audit",
"message": "fast-glob: Vulnerability found",
"severity": "high"
},
{
"type": "npm-audit",
"message": "micromatch: Vulnerability found",
"severity": "high"
},
{
"type": "npm-audit",
"message": "node-forge: node-forge RSA PKCS#1 v1.5 signature verification accepts extra nested DigestAlgorithm elements",
"severity": "high"
},
{
"type": "npm-audit",
"message": "shelljs: Vulnerability found",
"severity": "high"
},
{
"type": "npm-audit",
"message": "shx: Vulnerability found",
"severity": "high"
},
{
"type": "npm-audit",
"message": "source-map-js: source-map-js allows event-loop denial of service through indexed source-map section offsets",
"severity": "high"
}
],
"status": "WARNING",
"scannedAt": "2026-10-06T11:01:59.772Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}See how mu compares with popular alternatives.
Does this output matter? Is this command safe for the task? Is the work really done? A coding agent answers questions like these hundreds of times a session. Handing them to the main model or to fixed rules costs you in three ways:
Asking the main model each of these costs tokens and its attention, and fixed rules get them wrong too often. mu asks a judge: a small, fast model that answers one bounded question at a time, at 41 decision points in the agent loop. 34 of them act on their verdicts out of the box. Every verdict goes into a local ledger with its probabilities and timing (mu ledger, /status or the desktop app's Judge tab), and any decision point can run in shadow mode, asked and logged but changing nothing, so you can compare judges before you rely on one.
mu comes as a command-line tool, mu, which does everything pi does and adds the judgment kernel, and as a desktop app that bundles mu and its runtime. The default judge model is Jev, hosted by TypeSafe. It answers yes/no, choice and score questions with a probability for each answer. Any judge model can take its place: mude 1.0, mu's own model, on your Mac; another decision model from Hugging Face; your own System One server; or any LLM.
Contents: Quick start · How a turn works · Decision points · Judges · How mu is built · Measured · The hive · The board · Desktop app · Command line · Privacy · Documentation · Contributing
Desktop app. Download the installer for macOS, Windows or Linux from Releases, open it and sign in with ChatGPT, or paste an API key for any provider. There's nothing else to install.
Command line.
npm i -g mu-agent # Node.js 22.19 or newer
mu setup # connect a model
mu # start a session in the current directory
You don't need a judge key to start: until you set one, the free Jev on OpenCode Zen answers. Out of the box, 34 of the 41 decision points act on their verdicts, and every verdict is logged (/status, mu ledger or the app's Judge tab). To record what the judge would do without acting on it, put {"modes": {"default": "shadow"}} in ~/.mu/agent/mu.json. More in Getting started.
Each decision point is active (the default), shadow (asked and logged, but changes nothing; useful for comparing judges before you switch one on) or off. Each can also name its own judge: jev, mude (local), classifier:<provider>/<model>, llm:<provider>/<model> or several in order, such as mude,jev.
The 41 points fall into five groups:
/goal condition is met and what the board and the model should hear.| Decision point | Question | Effect |
|---|---|---|
input.preflight |
What kind of message is this, and how much thinking does it need? | A one-line hint to the model; optionally, the turn's thinking level |
task.frame |
Is this a new task, a hard constraint, a correction, a subgoal or no change? | Only a change rewrites the task frame: the goal, your constraints verbatim with their source, and the acceptance criteria |
input.interjection |
You send a message while the agent is working: interrupt now, after this step, or is it only asking how far along the work is? | The turn is cut short, or the message waits; a question about progress is answered by mu, and the agent isn't interrupted |
| Decision point | Question | Effect |
|---|---|---|
skills.disclosure |
Which skills are relevant to this task? | Only those enter the prompt; the rest can still be found |
capability.disclosure |
Does this task need an installed pack or MCP server? | Only then is it opened and its process started |
tool.admission |
For each chunk of a long tool output: does it matter now? | What matters enters the context; the rest is archived behind a pointer |
| `tool.admission.test-lo |
mu is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by qybaihe. mu (μ): most of what a coding agent decides isn't code. mu hands 40 kinds of yes/no calls to a small model: fewer wasted tokens, fewer approvals on autopilot, fewer runs that stop halfway. Built on pi. It has 611 GitHub stars.
mu returned warnings in SkillsLLM's automated security scan. It has no critical vulnerabilities, but review the flagged issues in the Security Report section before adding it to your workflow.
Clone the repository with "git clone https://github.com/qybaihe/mu" and add it to your Claude Code skills directory (see the Installation section above).
mu is primarily written in TypeScript. It is open-source under qybaihe on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh mu against similar tools.
No comments yet. Be the first to share your thoughts!