Description
Run a coding agent for an hour and most of your tokens go to decisions that have nothing to do with the code: is this command safe, does this long log belong in context, is that finding worth telling another agent. Leave them to the big model and it's slow and pricey; hard-code them as rules and they're wrong too often. mu hands them to a small, fast judge model called Jev so the big model keeps its attention for the actual work.
Built on the pi terminal agent framework, it asks Jev one bounded question at each of 30-plus decision points per turn, for example which chunks of a test log matter right now: failures stay, noise gets archived behind a pointer. Every verdict lands in a ledger you can review later. There's a desktop app and a CLI; install, connect a model, start working.
Judgment kernel: input classification, command risk, tool-output admission, context forgetting and compaction and more are scored by Jev. The answers only steer what the model does next; they never skip a step that needs your approval.
Leaner context: long outputs are judged chunk by chunk. Useful parts stay, the rest is archived behind a pointer and can be fetched back, saving tokens and attention.
Swarm collaboration: multiple sub-agents research or execute in parallel, every message delivery between them is recorded, and a side panel shows who is thinking about what.
Desktop app: installers for macOS (signed and notarized), Windows and Linux. Paste an API key or sign in with a ChatGPT, Claude, Grok or Google subscription, no extra runtime to install.
Command line: npm i -g mu-agent, run mu setup to connect a model. Everything pi does, plus the judgment kernel.
Judgment ledger: every verdict is logged. Review it with /status, mu ledger or the desktop judgments tab, or switch to shadow mode to record without acting.
Open source: MIT licensed, still in early development and used daily by its authors.
Built on the pi terminal agent framework, it asks Jev one bounded question at each of 30-plus decision points per turn, for example which chunks of a test log matter right now: failures stay, noise gets archived behind a pointer. Every verdict lands in a ledger you can review later. There's a desktop app and a CLI; install, connect a model, start working.
Features
Judgment kernel: input classification, command risk, tool-output admission, context forgetting and compaction and more are scored by Jev. The answers only steer what the model does next; they never skip a step that needs your approval.
Leaner context: long outputs are judged chunk by chunk. Useful parts stay, the rest is archived behind a pointer and can be fetched back, saving tokens and attention.
Swarm collaboration: multiple sub-agents research or execute in parallel, every message delivery between them is recorded, and a side panel shows who is thinking about what.
Desktop app: installers for macOS (signed and notarized), Windows and Linux. Paste an API key or sign in with a ChatGPT, Claude, Grok or Google subscription, no extra runtime to install.
Command line: npm i -g mu-agent, run mu setup to connect a model. Everything pi does, plus the judgment kernel.
Judgment ledger: every verdict is logged. Review it with /status, mu ledger or the desktop judgments tab, or switch to shadow mode to record without acting.
Open source: MIT licensed, still in early development and used daily by its authors.
