Muse Code CLI cheat sheet: commands, flags, and shortcuts

Muse Code is a terminal program, and the muse command does almost everything. This page collects the commands, flags, slash commands, and shortcuts from Meta’s documentation as of CLI 1.4.13. Type /help in a session for the full list your version supports, and /keymap for the shortcuts.

Install, sign in, and check your version

Command	What it does
curl -fsSL https://dev.meta.ai/install.sh | sh	Install on macOS or Linux
irm https://dev.meta.ai/install.ps1 | iex	Install on Windows (PowerShell)
muse --version	Show the installed version
muse	Start a session in the current directory; signs you in on first run
muse auth set	Store an API key instead of using the browser
/login	Reopen the sign-in options from inside a session
muse logout	Clear the stored browser session and key
muse usage	Show how much of your plan’s allowance is left
/upgrade	Open Accounts Center to change your plan

Muse Code updates itself. Our install guide covers updating, troubleshooting, and uninstalling.

Session slash commands

Command	What it does
/help	List every command available in this workspace
/new	Start a fresh session, keeping terminal scrollback
/clear	Start a fresh session and wipe scrollback
/resume	Pick an earlier session in this workspace to continue
/resume --last	Continue the most recent session
/name [name]	Show or set the session name
/fork	Branch a new session from the latest stable point
/rewind	Branch from an earlier message of yours
/side [prompt] or /btw <prompt>	Open a side conversation without derailing the main one
/compact	Summarize earlier context to free room
/recap	Summarize recent activity
/copy	Copy the last response
/export transcript	Save the conversation as a .txt file
/export trajectory	Save the full session log as JSON

Planning, goals, and agents

Command	What it does
/plan	Produce a plan and wait for your approval before building
/grill	Stress-test the plan until it holds up
/goal <objective>	Set a goal the agent works toward; /goal pause, /goal resume, /goal edit, /goal clear manage it
/loop <cadence> <prompt>	Repeat a prompt on an interval (10m) or a cron expression
/workflows	Open the control room for multi-agent workflow runs
/subagents	Show running and finished subagents
/tasks	Open the background tasks drawer
/stop	Cancel background tasks and terminals

The bundled skills post explains how /plan, /grill, and /goal fit together.

Models and reasoning

Command or flag	What it does
/model <model-id>	Switch model directly, e.g. /model muse-wonder-1.0
/models	Open the model picker
/effort	Set reasoning effort for the session
--model <id>	Choose the model at launch
--reasoning-effort <level>	none, minimal, low, medium, high (default), xhigh, max, or ultra
/settings	Settings, including Model → Contributor mode on plans
--contributor-mode on|off	Override Contributor mode for one run

Reasoning tokens bill as output, so save max and ultra for hard problems. Setting effort to ultra can also start a multi-agent workflow on its own.

Keyboard shortcuts

Key	What it does
Enter while the agent works	Steer: add your text to the running turn
Alt+Enter	Queue text for the next turn
Esc	Interrupt the current turn
Esc on an empty prompt	Take back the last queued instruction
Esc Esc when idle	Rewind
Ctrl+C	Interrupt (empty prompt) or clear the input
Alt+V (Option+V on Mac)	Start or stop voice input (not on Windows)

When the agent asks clarifying questions, number keys pick options, Space toggles multi-select, Tab adds a note, and Enter submits.

Running headless: muse exec

Flag	What it does
muse exec "<prompt>"	Run one task to completion and exit
--json	Print JSONL events to stdout
--prompt-file <path>	Read the prompt from a file
--max-model-steps <n>	Cap how much work the run can do
--session-id <uuid>	Continue an earlier session non-interactively
--no-session-log	Don’t keep a session log
muse export --session <uuid> --out run.json	Export a session

Exit codes: 0 the turn completed, 1 it failed or was cancelled (including hitting the step limit), 2 usage error, 130/143 interrupted. A 0 doesn’t mean the code is right — run your tests. In CI, authenticate with META_API_KEY; our API key guide has an example.

Permissions and the sandbox

Flag	What it does
--approval-mode on-request	Default: risky shell commands stop for review
--approval-mode untrusted	Anything without an allow rule stops for review
--approval-mode never	No approval prompts; the sandbox still applies
--approval-judge off	Send every approval decision to you, not the built-in judge
--disable-approval	Skip approvals, keep the sandbox
--disable-sandbox	Keep approvals, drop the sandbox
--yolo	No approvals, no sandbox, workspace trusted — disposable containers only
--sandbox-network <mode>	proxy-only (default), restricted, or unrestricted
--trust-workspace	Trust the folder without the first-run prompt

Sessions start in the Auto-review profile; the others are Ask me, Read-only, and Unrestricted. The docs explain what each one allows.

Skills, MCP, and project setup

Command	What it does
muse init	Write an AGENTS.md for the project (--dry-run to preview)
muse skills list	List installed skills
muse skills install ./my-skill --scope user	Install a skill
muse skills import --from claude	Import Claude Code skills (--from codex for Codex)
/migrate	Import Claude Code or Codex memory and MCP config
/mcp	Show connected MCP servers and tools
muse mcp login <server>	Sign in to a remote MCP server
muse config status	Show which settings are in effect
/theme	Switch the terminal theme

Where things live

• Settings: ~/.config/muse/settings.json — default permission profile, MCP servers, hooks, session retention.
• Project instructions: AGENTS.md; Muse Code also reads CLAUDE.md, .agents/AGENTS.md, and .claude/CLAUDE.md.
• Project skills and memory: .agents/skills/ and .agents/memory/ in the repository.
• Project hooks: .muse/hooks.json, loaded only after you trust the folder.

New to Muse Code? Start with getting started, then keep this page open in a tab.