Run Muse in a tab

Agents: muse · On: phone, desktop

A tab can run Muse, Meta's Muse Code CLI (muse). VCode drives it through muse serve, so its answers, edits, plans, todos and approvals show in the same transcript as the other agents.

Why you would use it

You have a Muse Code account and want to use it from the phone, with the same controls you use on a Claude, Codex, OpenCode, Cursor, Devin or Copilot tab. This page lists what Muse does in a tab and what it does not.

How to use it

Install the CLI if the workstation does not have it. See install-an-agent.
Log in to Muse. Muse Code needs a Meta account with a payment method. See log-in-to-an-agent.
Start a tab with Muse as the agent, in the new-thread card or with /new. Pick a model in the agent's menu. The list is the one Muse reports.
Set the effort on the same menu, when the model has efforts. A model's efforts are the variants Muse lists for it.

The agent menu with Muse in its list under Claude Code, Codex, OpenCode, Cursor, Devin and Copilot, with its mark and the line saying what picking it does.

What works

Steering. A message you send while a turn runs joins that turn. See queue-or-steer.
Stop. Stop ends the running turn. See stop-a-running-turn.
Model and effort. A change applies to the next message. Muse is not restarted.
Resume after a restart. The tab picks its Muse session up again when VCode restarts. See resume-after-a-restart.
Adopting a session. A Muse session on the workstation is listed on the resume screen. Adopting it opens a tab with the session's conversation in it. See resume-a-session-from-disk.
Fork. A fork opens a new tab with a copy of the conversation as text. See the limits below.
Side questions. /btw asks a side question. See side-questions.
/compact. VCode runs the compaction itself with Muse's own compact call. You see compacting: context at 120k, then compacted: context was at 120k.
Approvals. Below Full, a request from Muse shows as a card with Allow, Allow for session and Deny. See permission-levels. If Muse refuses your answer, a row reads Muse did not take that decision (<Muse's error>); decide again and the card shows again.
Plan. Plan turns on Muse's read-only profile. The plan Muse writes becomes the plan card; tap Build on it to turn Plan off and let Muse build.
Todos. The todo list Muse keeps shows in the plan strip and follows each change it makes.
Subagents. A subagent Muse starts shows in the subagent strip.
Images. Images you attach go to Muse inline with the message.
Usage meter. Once Muse reports usage, the usage meters show a Muse card with a meter for its session window and one for its weekly window. Before the first report, and on an account Muse never reports usage for, there is no Muse card.
Context meter. The tab's context figure follows the size Muse reports.
/mcp. The sheet lists Muse's MCP servers. You can log a server in and out there. See mcp-servers.

The levels

VCode	What Muse does
Ask	Asks before it runs a request. Each request is a card.
Auto	Reviews each step itself and asks you only when its review says to.
Full	Asks nothing.
Plan (flag)	Muse's read-only profile, on any level. It only writes its plan.

Changing the level restarts Muse and resumes the session with the new level.

What you see

Ask, Auto and Plan need Muse's sandbox. Where Muse reports that its sandbox cannot run, the tab does not start. On Windows it shows Muse's sandbox is not set up on this computer, so Muse runs at Full only. Run `muse sandbox windows setup` at the computer, or set this tab to Full. On Linux and macOS it shows Muse's sandbox cannot run on this computer (bubblewrap is not installed), so Muse runs at Full only. Set this tab to Full. The reason in brackets is the one Muse gives: user namespaces are turned off or bubblewrap is not installed. With no reason from Muse, the line has no brackets. Plan refuses at Full too, for the same reason.
An account with no payment method ends its login with Muse Code requires a payment method — subscribe at https://accountscenter.meta.com/muse_code/?ep=no_payg, then log in again. The card shows that line.
A login file VCode cannot read as a JSON object shows could not read Muse's login.
Without a login the agents card shows not logged in and a Log in button.
Each finished edit is one row: edited 3 file(s), +12 −4.
When Muse refuses /compact, the row reads compact failed: Muse refused the compaction (<Muse's reason>).

Options and settings

Setting	Default	What it changes
`META_API_KEY` (environment variable on the workstation)	unset	An API key that counts as a login. The agents card names it `API key`, and it wins over a saved login.
`MUSE_AUTH_PATH` (environment variable on the workstation)	unset	The login file VCode checks. Without it VCode checks `~/.config/muse/auth.json`, or the same path under `XDG_CONFIG_HOME`.
Update every hour, `agentUpdate.muse`	off	Runs `muse --version` with `MUSE_SYNC_UPDATE=1` once an hour when no Muse tab works. See update-agents-automatically.

Limits and known gaps

Not in this version. Rewind, goals, workflows, computer use, plugins and recaps are not offered on a Muse tab.
Edits show as counts. A row gives the number of files and the lines added and removed. VCode shows no diff for a Muse edit.
Fork and side questions copy text. Muse's own fork does not work on Linux, so a fork gives the new session the conversation as text. The copy is a new session, not the old one. A side question is a read-only one-shot run that gets the conversation as text.
A logged-out Muse lists no models. The model menu is empty until you log in.
MCP servers: log in and log out only. A Muse row offers Authenticate when it needs a login and Log out when it is logged in. There is no reconnect, no turn off or on, and no remove, because Muse's protocol cannot change its servers. A server added for one session only offers neither.
Adoption imports the last 200 items of the session at most. The agent still has the whole session.
The resume list is at most a minute old. VCode reads Muse's session list at most once a minute. A list that takes longer than 10 seconds keeps the rows it had.
Instructions ride in the first message. VCode's instructions go to Muse inside a <vcode-system> block at the top of the first message of a new session. See what-the-agent-is-told.
Binary lookup. On Linux and macOS VCode uses muse on PATH, then ~/.local/bin/muse. If neither exists the tab fails with muse is not installed (looked on PATH and in ~/.local/bin). On Windows it looks for the newest build in %LOCALAPPDATA%\Programs\muse, and fails with muse is not installed (looked in %LOCALAPPDATA%\Programs\muse).

the-three-agents: the agents and what each one can do
copilot: the GitHub Copilot tab
devin: the Devin tab
permission-levels: the levels and Plan
switch-the-model: the model and effort menu
resume-a-session-from-disk: finding and adopting sessions
install-an-agent: installing the CLI from the phone
log-in-to-an-agent: the login that Muse runs
mcp-servers: /mcp on a Muse tab, with its two operations
update-agents-automatically: the hourly update check