Last updated: September 28, 2026

How to Use Muse Code: From Install to First Task

1. Install the CLI. macOS/Linux: curl -fsSL https://dev.meta.ai/install.sh | sh. Windows (PowerShell): irm https://dev.meta.ai/install.ps1 | iex.
2. Check the install. Run muse --version.
3. Start a session in your project. Run muse inside your project directory and approve the workspace trust prompt.
4. Sign in. Choose browser sign-in, or paste a Meta API key. For scripts and CI, set the META_API_KEY environment variable.
5. Give it your first task. Type a task in the composer. Press Enter to steer a running turn, Alt+Enter to queue for the next turn, Esc to cancel.

How to install Muse Code (Mac, Linux, Windows)

macOS and Linux

curl -fsSL https://dev.meta.ai/install.sh | sh

Windows (PowerShell)

irm https://dev.meta.ai/install.ps1 | iex

Muse Code runs natively on Windows; you do not need WSL. Then check the install:

muse --version

Only install from dev.meta.ai. Third-party "Muse Code download" sites are not official.

Source: Official docs

Sign in: browser or API key

The first time you run muse in a project, it asks you to sign in. There are two ways:

• Browser sign-in: a browser opens, you approve the session, and you return to the terminal.
• API key: paste a Meta API key instead. For scripts and CI, export it:

export META_API_KEY="<your-key>"

You can also store a key with muse auth set. The environment variable wins over a stored key, which wins over a browser session. Meta Managed Account (MMA) users must use an API key. Sign out with muse logout.

Source: Official docs

The workspace trust prompt and permissions

On first launch in a folder, Muse Code asks you to trust the workspace. Project instructions (like AGENTS.md) only load in trusted workspaces. After that, two safety layers are on by default:

• Approvals: safe actions run automatically; risky ones (for example a destructive rm -rf) wait for you. You can allow once, always allow in this workspace, or reject.
• Sandbox: shell commands run inside an OS sandbox (Seatbelt on macOS, bubblewrap on Linux, Windows sandbox on Windows), limited to your workspace and temp folders. .git, .muse and .agents stay read-only.

Network access defaults to proxy-only (you approve each destination); use --sandbox-network restricted for none. Use --yolo only in an isolated environment.

Source: Official docs

Run your first task

Type a task and press Enter. You don't have to wait for the agent to finish: press Enter to add instructions to the running turn, Alt+Enter to queue them for the next turn, and Esc to cancel the turn. You can paste images into the composer and refer to them as [Image #N]; Alt+V toggles voice dictation (not on Windows).

Source: Official docs

Muse Code commands: cheat sheet

Source: Official docs

Project rules: AGENTS.md (and CLAUDE.md)

Create a project instruction file with:

muse init            # creates AGENTS.md
muse init --dry-run  # preview first

Muse Code looks for AGENTS.md, CLAUDE.md, .agents/AGENTS.md and .claude/CLAUDE.md, from the workspace root up to the nearest .git. So if you already use Claude Code, your existing CLAUDE.md works. User settings live in ~/.config/muse/settings.json.

Source: Official docs

Choose the model and reasoning effort

Switch models mid-session with /models, or at launch with --model. Reasoning effort goes from none to ultra (default high); change it with /effort or --reasoning-effort. Higher effort uses more output tokens, which matters if you pay per token (see Muse Code pricing).

Source: Official docs

Multi-agent: subagents and workflows

A lead agent can hand bounded tasks to subagents, optionally each in its own git worktree so parallel edits don't collide. One agent tree runs up to 8 agents at once by default, configurable up to 64. For bigger jobs (parallel audits, comparing approaches, staged migrations) Muse Code can run a workflow: it orchestrates child agents and returns one combined result. Keep small, tightly coupled edits on the main agent.

Source: Official docs

Undo a wrong turn: rewind

Run /rewind or press Esc twice on an empty composer, pick an earlier message, and Muse Code opens a new branch from there. The original session is kept. Rewind changes conversation history only: it does not revert files in your workspace, so use git for that.

Source: Official docs

Run Muse Code in scripts and CI

muse exec "run the tests and fix any failures" --json

Headless mode has no UI. It exits 0 on success, 1 on failure and 2 on usage errors. Cap runs with --max-model-steps, and authenticate with META_API_KEY.

Source: Official docs

Muse Code on Windows: what's different

• Commands run in PowerShell instead of Bash.
• The sandbox may ask for administrator approval the first time.
• Voice input and session messaging are not available on Windows.

Source: Official docs

Using Muse Code in VS Code

The official docs don't list a VS Code extension. Open VS Code's integrated terminal and run muse there. Treat any "Muse Code for VS Code" extension as unofficial.

Limits to know before you start

• Plans cap usage in 5-hour windows (Everyday Usage: 10-50 prompts every 5 hours). When you hit the limit you upgrade or wait for it to refresh.
• The cheap Contributor token tier allows Meta to train on your prompts and completions, and has a lower rate limit.
• Hooks and MCP servers run outside the sandbox and approvals.

Costs and plans are on the Muse Code pricing page. Next, read our Muse Code review or compare Muse Code vs Claude Code.