mu

mu is a coding agent built on the idea that most of what an agent decides is not code. Does this output matter? Is this command safe for the task? Is the work really done? It sends those bounded questions to a small judge model at 40 decision points in the loop, instead of the main model or fixed rules.

Why it matters: the judge reads long output in chunks so a 2,000-line log does not bloat every later request, approves what the task clearly needs and asks about the rest, and catches a run that stopped halfway or a "done" nothing verified. Every verdict is logged. The default judge is Jev.

TypeScript, MIT, about 570 stars at the time of listing. A desktop app, or npm i -g mu-agent for the CLI.