mu
Coding agent with a judgment kernel: a small judge answers about 35 bounded calls per turn while the big model does the work. CLI and desktop, built on pi and AionUi.
Overview
mu (github.com/qybaihe/mu, npm mu-agent) is a coding agent with a judgment kernel, built on pi and AionUi. A small judge answers one bounded question at a time at 35 decision points in every turn. The big model keeps the work. Each point is active, shadow, or off, and can name its own judge: hosted Jev through TypeSafe, OpenRouter, or the Vercel AI Gateway, local Laya (322M, on the machine, nothing downloaded without consent), an llm: provider tier, or a cascade such as laya,jev. Shadow asks and logs and changes nothing, so you can compare judges before a point is allowed to act. The CLI is npm i -g mu-agent on Node 22.19 or newer (mu, mu doctor, mu -p, mu ledger). Desktop builds for macOS (Apple silicon and Intel), Windows (x64 and Arm), and Linux (x64 and Arm) ship from GitHub Releases with the runtime inside. README calls 0.1.x early pre-releases whose names and formats may still change. kyrn/npm/CHANGELOG.md records 0.1.7 on 2026-09-30. Root LICENSE covers packages and kyrn (MIT). desktop/ keeps the AionUi Apache-2.0 license. Public counts were 336 stars and 32 forks when this listing was drafted.
Problem: Coding sessions spend the big model on calls that are not the code: what stays in context, whether a flagged command was asked for, whether a finding belongs on a shared board, and whether done was actually verified.
Built for: Developers who want a full coding agent, CLI and desktop, where a small judge owns about 35 bounded decision points and the big model still writes the code.
First indexed on Jev Directory: 2026-10-01
Creator and team
- Name
- qybaihe
- Handle
- @qybaihe
“mu gives them to a judge: a small, fast model that answers one bounded question at a time, at 35 decision points in every turn.”
How Jev is used
- Role in the product flow
- Judgment kernel inside the agent turn: a small judge answers bounded questions at 35 decision points while the big model does the work
- Primitives
- ChoiceScoreNoul
- State in
- A short question and a small state per decision point: message kind, task frame, tool chunk, flagged command, bee finding, or board line. README Privacy: the judge sees only the fields that question needs.
- Decision out
- A verdict with a probability, written to the local ledger (mu ledger, or the desktop judgments tab). Active points change the turn. Shadow points are logged and change nothing. Off points are not asked.
- Input: input.preflight, task.frame, and input.interjection before the model starts
- Model calls a tool. Rules catch a dangerous-looking command first. tool.risk, tool.constraint, and tool.approval run in Jev approves mode
- tool.admission walks long output chunk by chunk. context.forget and context.compact run when context grows. Test-log folding is rules by default
- Turn end: turn.completion, turn.drift, turn.rewind, memory.applied, board.read, and cache.warming
- Hive: hive.publish, hive.deliver, and hive.relate on a shared append-only board. Bees read and report. The main model edits
mu is the agent: CLI, desktop, and a judgment kernel inside the turn. y0usaf-pi-jev is a Pi extension that batches gate nouls on bash, write, and edit. devmortimer-pi-warden steers a Pi session with guardrails. leepokai-jev-guard risk-scores tool calls and scans results across several hosts. samuelfaj-distill keeps Jev on consult, effort, and compression inside a Rust TUI, and Distill owns permissions. tianyucodings-jevharness lets an LLM write a harness and then calls Jev for fast task decisions. kitfunso-hippo-memory reranks local memory rows. kerpopule-hermes-jev-skills is a Hermes skill drawer for routing, memory, compaction, and GUI steps. mu owns the session. About 35 points run from input.preflight through hive.relate. board.read feeds a plain-language board written by a model picked only because it speaks plainly. Bees (2 to 6) read code, run commands, and browse. They do not edit. The main model makes the change. Rules still catch a dangerous-looking command first. The judge only vouches that you asked for it, and an unsure tool.risk verdict asks you. In Jev approves mode, tool.approval runs only what it is sure the task needs. README is explicit about what the test-log numbers are not: they measure what reaches the model, the goals and labels are the authors', the repeat chart is 7 Vitest logs from one developer, and there is no end-to-end comparison with pi, Claude Code, or Codex yet. Cite github.com/qybaihe/mu README, kyrn/docs/09-test-log-admission.md, and kyrn/npm/CHANGELOG.md.
Sourced performance claims
- README lists 35 decision points grouped as input, context, tools and safety, turn, and teamwork. Each point is active, shadow, or off, and can name jev, laya, llm:provider/model, or a cascade.Source: github.com/qybaihe/mu README Decision points
- Authors' sessions: one warm Jev question about 0.3 s over HTTP/2. 16 chunks of tool output judged in one request in 0.44 s, state billed once. Local Laya is 322M and does not touch the network.Source: github.com/qybaihe/mu README Judges
- Test-log replay, measured 2026-09-21: 7 failing Vitest logs, 139,820 characters, exact-repeat folding removed 51% with no model call. Goal-aware Jev selection cut 40.2% on 29 tuned goals and lost 0 of 72 required lines, and cut 46.4% on 9 held-out goals and lost 0 of 19. 282 live Jev requests cost about $0.017 at list price. Median request 345 ms. Both arms are off by default.Source: github.com/qybaihe/mu README Measured and kyrn/docs/09-test-log-admission.md
- Hive example in the README: 3 bees, 9 minutes, 117 candidates judged, 27 notes on the board, 16 delivered to the bee that needed them. Bees do not edit.Source: github.com/qybaihe/mu README The hive
- Public GitHub repo qybaihe/mu had 336 stars and 32 forks on 2026-10-01. Last push 2026-09-30. kyrn/npm/CHANGELOG.md version 0.1.7 is dated 2026-09-30.Source: GitHub API 2026-10-01 and kyrn/npm/CHANGELOG.md
Features and stack
Features
- 35 decision points with per-point judge and active, shadow, or off
- Hosted Jev, local Laya, or an LLM tier, including cascades
- Test-log exact-repeat folding and optional goal-aware Jev selection
- Hive board with publish, deliver, and relate, plus a plain-language board
- Desktop app for macOS, Windows, and Linux with the runtime bundled
- CLI ledger, task frame, permissions, and import from Claude Code or Codex
Stack
- TypeScript
- npm mu-agent
- pi (MIT)
- AionUi desktop (Apache-2.0)
- TypeSafe Jev, OpenRouter, or Vercel AI Gateway
- Local Laya 322M
Pricing: MIT packages and an Apache-2.0 desktop shell. You pay the big-model provider plus any hosted Jev calls. Local Laya stays on the machine. README says the 282 Jev requests in the test-log study cost about $0.017 at list price.
Links
FAQ
- Is mu the same as the pi-jev extension?
- No. y0usaf-pi-jev is a Pi extension that gates bash, write, and edit. mu is a coding agent built on pi, with its own CLI, desktop app, and about 35 decision points inside the turn.
- What do active, shadow, and off mean?
- README: active changes what the model does next. Shadow asks the judge and logs the verdict and changes nothing, which is how you compare judges before switching a point on. Off does not ask. /mu mode sets one point.
- Do hive bees edit the repo?
- README: a hive is 2 to 6 bees. They read code, run commands, and browse. They never edit. The main model makes the change. hive.publish asks whether a finding is worth the shared board.
- What do the test-log savings measure?
- README Measured: exact-repeat folding is rules, no model call, and removed 51% of characters across 7 logs totaling 139,820 characters. Goal-aware Jev selection is a separate arm, off by default, and the README says the numbers do not show whether the model then finishes the task.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.