Muse-Desktop

<div align="center">

# Muse-Desktop

**The desktop app Muse Code was missing.**

[![Release](https://img.shields.io/github/v/release/EtienneLescot/muse-desktop?style=flat-square&color=0b7285&label=release)](https://github.com/EtienneLescot/muse-desktop/releases/latest)
[![Platforms](https://img.shields.io/badge/platforms-macOS%20%C2%B7%20Windows-495057?style=flat-square)](#install)
[![License](https://img.shields.io/badge/license-MIT-495057?style=flat-square)](LICENSE)
[![Built with Tauri](https://img.shields.io/badge/built%20with-Tauri%20%C2%B7%20React-495057?style=flat-square)](https://tauri.app)

</div>

![Muse-Desktop on Windows](assets/screenshot-windows.png)

Muse Code is Meta's coding agent, and it ships as a CLI. Muse-Desktop is the harness that
goes with it: the kind of client Claude Code and Codex users take for granted, built for
Muse. Projects, git worktrees, subagents, MCP, a diff review panel. And an agent that can
see your screen and use your applications.

> **0.1.0 · macOS (Apple Silicon) and Windows x64.**
> [Download the latest release](https://github.com/EtienneLescot/muse-desktop/releases/latest).
> The macOS DMG is signed and notarised. The Windows installer is not signed yet.

## What it does

**Let the agent use your computer.** Turn it on and Muse can look at your screen and work
in your applications: open a file, fill a form, click through a tool that has no API.
Two levels of access, and it starts at neither:

| Level | What Muse may do |
|---|---|
| **Off** | Nothing. This is the default. |
| **Observe only** | Take screenshots, read windows. No clicks, no typing. |
| **Observe and act** | Mouse, keyboard, clipboard. |

You choose the level, you can revoke it at any moment, and the permission is a list of
named tools rather than a blanket "allow". It runs on the open-source
[cua driver](https://github.com/trycua/cua) (MIT), installed separately. Settings walks you
through it.

**Work by project.** A project is a folder. Conversations are grouped under the folder they
run in, reorderable by drag, and each one carries its own model, reasoning effort and
permissions. They survive a restart.

**See what the agent is doing.** Thinking shows up before the first word of the answer, in
a block you can fold away. Tool calls stream live, then collapse into one line when they
are done. Sub-agents get their own lane. Answers are Markdown, wi