v0.3.0Fewer approval prompts, fewer wasted tokens, fewer runs that stop halfway · mu-agent 0.3.0

mude 1.0: a decision model optimized for coding

mu's own decision model, trained for the decisions a coding agent makes and state of the art at making them. It runs on your machine, so what it judges never leaves it.

A judge that runs on your machine

2B parameters

One forward pass answers every question in a decision, each with a probability.

Judged locally

Its server runs on 127.0.0.1. What it judges never leaves your machine.

MIT license

Both editions are on Hugging Face, free to download.

On Hugging Face

One command on a Mac

One-click install works on Macs with Apple silicon today, with macOS 14 or later and at least 8 GB of memory. On such a Mac, mude is the recommended judge, and mu setup offers it first. While it runs, the server uses about 4 GB.

mu judge mude install               # download (resumes after a stop), verify every file, start it on 127.0.0.1, make it the judge
mu judge mude status                # what is installed, running and enabled
mu judge mude enable --local-only   # mude only: nothing goes to a hosted judge
mu judge mude disable               # back to the previous judge; stops its server
mu judge mude uninstall             # remove the model and its Python environment

In the desktop app

First-run guideOn a Mac that can run it, the judge step starts with mude selected as the recommended judge, and one click installs it. If an install stops, click Continue installing to pick up where it left off.
Settings → JudgesThe mude card installs, starts and stops it, makes it the judge and uninstalls it. It also has the Local only switch.

On other machines

Elsewhere, run the full edition (mude-1.0 on Hugging Face) with its own System One server, as its model card shows, and add it to ~/.mu/agent/mu.json as a typesafe judge. MUDE_KEY holds the key the server was started with:

{
  "judges": {
    "mude": {
      "type": "typesafe",
      "baseUrl": "http://127.0.0.1:8790/v1/systemone",
      "model": "mude",
      "apiKeyEnv": "MUDE_KEY"
    }
  },
  "tiers": ["mude"]
}
Using mude 1.0 with mu, on its model card

What stays on your machine

With a fallback (the default)

The judge you used before stays on as mude's fallback. It answers what mude can't (when mude's server isn't running or mude is unsure) and always handles the two injection screens: tool descriptions and tool output. When that judge is hosted, what goes to it leaves your machine.

mu judge mude install

Local only

mude is the only judge, and nothing goes to a hosted judge. When mude can't answer, the decision point does what it would do without a judge.

mu judge mude enable --local-only

mude 1.0 and Jev

mude 1.0Jev
Where it runsOn your machine, at 127.0.0.1Hosted by TypeSafe, reached through TypeSafe, OpenRouter, the Vercel AI Gateway, OpenCode Zen or Cloudflare Workers AI
What leaves your machineNothing, with Local only on. Otherwise, what goes to the fallback judgeThe fields each question needs, such as a message, a chunk of output or a command. Credentials are masked first
PlatformsOne-click install on Macs with Apple silicon. Elsewhere, the full model on your own serverWherever mu runs
License and availabilityMIT. Both editions are on Hugging FaceA hosted service: free with no key on OpenCode Zen for a limited time, or paid per call with your own key
Setupmu judge mude install, or one click in the desktop appNone, it's the default. mu setup sets a key

Capability leaderboard

See the mubench 1.0 leaderboard

If mu is useful to you, star it on GitHub

Star on GitHub553