mu runs a coding agent's routine decisions on a small judge model

A coding agent that writes real code also spends most of its session answering small questions that have nothing to do with the code: whether this file belongs in context, whether that shell command is safe, whether a finding is worth telling another agent, when the work is actually done. mu, an open-source agent by qybaihe, hands those routine calls to a small decision model and keeps the big model for the writing. The repository, opened on 22 September 2026, reached 524 stars on 56 forks.

What ships

mu is MIT and dual Apache-2.0 licensed, packaged as both a desktop app for macOS, Windows and Linux and a CLI (mu-agent on npm, requires Node.js 22.19 or later). It is built on pi, an open-source coding agent, and on AionUi, a desktop framework. The default judge is OpenAI's Jev-1.13-free; users can swap in a local decision model called Laya, or any large language model, and the big worker model is selected per session from Claude, ChatGPT, Grok or Google through in-app login. The repository lists 38 configurable decision points per turn where a judge call runs in place of a full worker call.

What the author measured

The README publishes three numbers from the author's own testing. Folding exact repeats in failing test logs saved 51 percent of characters without losing data. On 29 goals that the author tuned by hand, the judge cut 40.2 percent of a test's output while keeping every line marked as required. One warm judge question took around 0.3 seconds; 16 tool-output chunks were judged together in 0.44 seconds. These are the author's own figures from a project the README describes as "early development, pre-releases 0.1.x" and actively used in daily work by its authors, so a buyer should re-run them before trusting any one of them in production.

Where the design fits

The pattern mu names, a small fast judge in front of a larger worker, is close to what several other projects on GitHub are now shipping, but mu's claim is specifically about the token cost of a coding agent rather than about model routing between providers. A read of a long test log, a check that a command is in a safe-list, a judgement of whether a diff is ready for review, a choice of which three files to keep in the next turn: each of these is a bounded yes-or-no or short-scoring question that a 1-billion-parameter decision model can answer in milliseconds. The worker model never sees the raw text and never pays for it in input tokens.

For a team already running a coding agent on their own developers' machines, the trade here is more plumbing in return for a smaller bill. The judge is a second model to install, keep updated and watch for drift, and a cut log is a log a human cannot read without going back to the original. The 40.2 percent cut is a number from one hand-tuned test set, so the first useful experiment is to run mu beside a current agent on the team's own repository for a week and compare what the big model actually billed for.

Source

qybaihe/mu on GitHub.