v0.3.0Fewer approval prompts, fewer wasted tokens, fewer runs that stop halfway · mu-agent 0.3.0
mu
English
Star 553
Install

Need a decision model customized for you and your company?

In every turn, mu asks a decision model dozens of small questions: may this command run, which old output is worth keeping, is the work done. We can tune one to how your team answers them, run it on your own machines, and show you in numbers what it changes.

Tell us what you needAbout mude 1.0

Start from mude 1.0, on your own machines

Macs with Apple silicon

One click in the desktop app, or one command: mu judge mude install. mu downloads the 8-bit MLX edition, checks every file, starts it on 127.0.0.1 and makes it the judge.

huggingface.co/qybaihe/mude-1.0-mlx-8bit

Your GPU servers

The full bf16 model, with its own System One server. A CUDA GPU is recommended. mu reaches it as a judge over your own network.

huggingface.co/qybaihe/mude-1.0
More about mude 1.0 on the Models page

Tuned to how your team decides

  1. Your decisions

    We start from the decisions that matter to you: what counts as risky in your codebase, what your review policy asks for, how work should be routed.

  2. A baseline

    We score your current setup on those decisions the way mubench does: each answer by the cost of its consequences.

  3. Fine-tuning

    We fine-tune a decision model on your decision points, workflows and rules.

  4. Before and after

    The same scoring again, next to your current setup, so you can see what changed before you switch.

How mubench scores decision models

Works with mu out of the box

The judge in the CLI and the appmu judge mude install on the command line, or the mude card in the desktop app's Settings → Judges.
Local onlyWith mu judge mude enable --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.
Shadow mode firstDecision points in shadow mode still ask and record their verdicts, but change nothing, so you can watch a model before you trust it.
A ledger for auditEvery verdict is recorded on your machine: the answers and their probabilities, how long it took, and whether the judge or a fallback gave it. Read it with mu ledger or in the app's Judge tab.

Common questions

Can we try mude 1.0 before we talk?Yes. It's free under the MIT license. Install it, put your decision points in shadow mode, and read its verdicts in the ledger.
Does our coding model change?No. A decision model only answers mu's decisions. Your coding model, and where it runs, stay your choice.
What leaves our network?Nothing mude answers: it runs on your machines. With Local only on, nothing goes to a hosted judge either. Your coding model is separate, and what it sees goes wherever it runs.
How would we know a custom model is better?We score it the way mubench does, on your own decision points, before and after, next to your current setup. You see the numbers before you switch.

Tell us what you'd like a model for

Coming soonThis form isn't connected yet. Pressing Send sends nothing and stores nothing.

If mu is useful to you, star it on GitHub

Star on GitHub553