mu

Purpose

μ is a programming agent with a newly added "judgment kernel" for handling the many decisions in a session that are unrelated to the code itself, such as what stays in context, whether a command is safe, whether a finding is worth telling other agents about, whether a task is complete, and so on. The idea is to let a small, fast judgment model take on routine judgment, thereby freeing up the large model's attention for coding work.

Main capabilities

• Decision points: each turn has more than 30 decision points (the official text mentions 38), each of which asks the judgment model a short question; the answer only affects the model's next action and does not change whether it asks you a question. Each decision point can be set to active, shadow, or off, and can specify its own judge: jev, laya (local), classifier:<provider>/<model>, llm:<provider>/<model>, or a cascade such as laya,jev.
• Judge (Jev): accessed via TypeSafe, OpenRouter, Vercel AI Gateway, OpenCode Zen, Cloudflare Workers AI, or a compatible service, each requiring the corresponding key. When no key is set, the free Jev 1.13 on OpenCode Zen is used. Author's own testing: about 0.3 seconds for a hot question (HTTP/2); 0.44 seconds to judge 16 chunks of tool output in a single request.
• Local judge (Laya): 322M parameters, runs on the local machine, no network. Reliable for simple predicates, weaker at meta-judgments; recommended to first compare with Jev in shadow mode.
• Other judges: any classifier model in the pi directory, or any LLM.
• Test log trimming: supports exact duplicate collapsing (identical diffs, DOM dumps, or stack traces keep only the first instance, with the rest replaced by a one-line marker, without calling the model) and goal-aware selection (Jev judges whether the current goal needs each chunk of passing test output, discarding only when the probability of "not needed" reaches 0.9 or above). Both are off by default. In the author's research, 282 live Jev requests cost about $0.017 at list price, and the median request time under the default wording was 345 milliseconds. The official text explicitly states that these numbers measure what reaches the model and what is lost, and do not represent whether the model can complete the task; the duplicate data came from 15 Vitest logs from one developer over two days.
• Swarm (hive): 2–6 bees, each with a focus area, onlyread code, run commands, and browse; they do not edit, and the main model makes changes. After a bee speaks, hive.publish asks whether there is a finding, dead end, decision, or blocker worth sharing; the shared board is append-only, and hive.relate reads the relationship between two notes (supersedes, contradicts, supports); superseded conclusions are delivered as corrections, contradictory notes coexist, and a verification bee may be dispatched.
• Language panel (board): when enabled, the agent generates a plain one-line description on the panel at the end of each step; board.read asks the judge whether a given utterance is a new message, has a model specifically chosen to "speak plainly" rephrase it for a human, and maintains the panel state (what is currently happening, how much of the checklist is complete, what is waiting for you).
• Desktop app: bundles mu and its runtime, no Node installation required. Includes work panels for board, judgments (live ledger), hive, lessons, files, preview, source, browser, and more. Permission mode and goal are in the input area, and ⌘K opens the command palette. Supports ChatGPT, Claude, Grok, Google subscription login, or an API key for any provider supported by pi, and can import and continue Claude Code and Codex CLI conversations.

How to use

• Desktop app: download the macOS, Windows, or Linux installer from Releases, open it, and paste an API key or use subscription login. You can start without a judge key; the free Jev on OpenCode Zen is used by default.
• Command line: requires Node 22.19 or later.mu -p "prompt" run once and print;mu -c continue the previous session;mu import --list Find Claude Code and Codex conversations,mu import <file> import;mu ledger [n] print the judgment results of the most recent n sessions. The command line and desktop app share the account, settings, and lessons. pi's original commands (/model, /thinking, /login, /resume, /tree, /fork, /compact, /export, etc.) remain unchanged.MU_JUDGE=laya,jev mu can override the judge for a single run.
• Configuration: set shadow mode in ~/.mu/agent/mu.json {"modes": {"default": "shadow"}}, test log trimming "features": {"admission": {"testLog": "rules"}}.
• Desktop development: in desktop/ bun install, then KYRN_ROOT="$(cd .. && pwd)" bun run start(first in the root directory npm install).

Use cases

• Long-session programming tasks that need control over context length and token cost.
• Complex debugging or exploration that needs multi-agent division of labor, shared findings, and tracking of conclusion changes.
• Scenarios where you want to understand agent progress in plain language and reduce reading lengthy output yourself.
• Users who care about privacy and want the judgment process to stay on the local machine (using Laya or MU_JUDGE=off).

Limitations

• The project is in early development; 0.1.x is a pre-release version, and names, settings, and formats may change.
• No official end-to-end comparison with pi, Claude Code, or Codex on the same task is provided.
• The duplicate data in the test logs comes only from 15 Vitest logs within two days; other test runners were not measured.
• Laya is weak at meta-judgment; the accuracy and performance data come from the author's own sessions and replays, and it is recommended to compare in shadow mode before use.
• No official pricing or billing plan details are provided, only that Jev 1.13 is free on OpenCode Zen with a limit.
• Keys are kept on the local machine; mu will not download models or runtimes on its own, and anything that needs to be downloaded will be asked about first.