Run Meta's Muse Code on a remote server

Install Muse Code#

On your computer, create a VM and connect. Pick any free name; if you've already clicked Deploy or have a VM, skip the first command and use its name:

ssh exe.dev new --name=my-muse-code
ssh my-muse-code.exe.xyz

On the VM, install Muse Code with Meta's installer, make a project folder, and open a tmux session:

curl -fsSL https://dev.meta.ai/install.sh | bash
source ~/.bashrc
muse --version
mkdir -p ~/project && git -C ~/project init -q
tmux new -A -s muse -c ~/project

muse --version should print Muse Code 1.4.18 or newer. The installer puts a launcher at ~/.local/bin/muse and adds that folder to your PATH.

Sign in and start coding#

Inside the tmux session, start Muse:

muse

It asks whether to trust the folder (choose Trust and continue), then how to sign in. Sign in with your browser works over SSH: Muse prints a https://auth.meta.com/oauth/device/... link and a code. Open the link on your computer, check the code matches, and approve. You can also run muse login for the same flow. Muse keeps the sign-in in ~/.config/muse on the VM. A Muse Code subscription works only through this sign-in.

Then give it a task. For example:

Build a small command-line tool that prints the weather for a city, with tests.

For longer work, set a goal and let it run:

/goal Add a JSON output flag and make every test pass

Commands run in Muse's OS sandbox, which works on exe.dev VMs as it is: the agent can write in the project folder and its temp directory, and the rest of the disk is read-only to it. Risky commands and new network destinations need approval, and its automated reviewer clears the safe ones. See Permissions and safety for the other profiles.

Muse reads AGENTS.md or CLAUDE.md for project instructions (muse init writes a starter AGENTS.md). It also reads your personal Claude Code rules; on exe.dev those include notes on how the VM's HTTPS proxy works.

Leave it running#

Press Ctrl-b, then d to detach. Muse keeps working on the VM while you're gone. On your computer, use this command to hop back in:

ssh -t my-muse-code.exe.xyz 'tmux new -A -s muse -c ~/project'

When you come back after a couple of minutes away, Muse shows a recap of what happened. /tasks and /subagents show background work, and /stop stops background commands (Esc only interrupts the current turn).

A VM restart ends the tmux session but not your work. Reconnect and, from the project folder, run:

muse resume --last

Muse rebuilds the conversation from its log, notes that the last run ended abnormally, and continues. Meta's docs warn that an interrupted step with side effects is marked as unknown for the agent to check, not repeated blindly.

For a second, independent task, open another tmux window (Ctrl-b, then c) and start muse in another folder or git worktree. Named sessions can send each other handoffs with session messaging.

Make yourself at home#

• Bring your code: connect a GitHub repo and clone it with git clone https://github.int.exe.xyz/you/repo.git. No GitHub token goes on the VM. Muse won't commit or push unless you ask, and GH_HOST=github.int.exe.xyz lets it use gh for pull requests.
• Peek at your app: the sandbox's default network mode keeps ports the agent opens inside the sandbox, so a dev server it starts isn't reachable from your browser. Either start the server yourself in another tmux window, or start Muse with muse --sandbox-network enabled, which lets sandboxed commands use the network freely. Then serve on 0.0.0.0:8000 and open https://my-muse-code.exe.xyz/, which only you can reach. More about previews.
• Run it headless: muse exec "Run the tests and fix what fails" runs one prompt to the end, for scripts and cron. Add --max-model-steps to cap a run, and muse exec --session-id <uuid> "Continue." resumes one. See headless and CI.
• One VM per task: ssh exe.dev cp my-muse-code my-muse-code-2 copies the VM, with Muse, its sign-in, and your code, so a risky job can't touch the original. Or give an agent exe.dev's MCP server with a token limited to one tag, and it can make VMs of its own.
• Shelley is here too: exe.dev's own agent runs on the same VM at https://my-muse-code.shelley.exe.xyz/, and Claude Code and Codex are preinstalled, all on exe.dev's LLM integration with no key. Use one for a second opinion on what Muse built.

Other agents with pages here: Claude Code, Codex, Gemini CLI, Kimi Code, and the rest of the AI agents category.

Keep a Meta API key off the VM#

Pay-as-you-go accounts can use a Meta API key instead of signing in, and Meta Managed Account (Business & Enterprise) users must. muse auth set --api-key-stdin would store the key on the VM. To keep it at exe.dev's edge instead, put it in an HTTP proxy integration and point Muse at the integration.

On your computer, create the integration; paste the key at the prompt:

printf 'Meta API key: '; read -rs K; echo
printf %s "$K" | ssh exe.dev integrations add http-proxy --name meta-api \
  --target https://api.meta.ai --bearer=- --attach vm:my-muse-code
unset K

On the VM, point Muse's Meta provider at it and give Muse a placeholder key; the integration sends the real one to Meta. This adds one setting to ~/.config/muse/settings.json, which Muse creates on its first run:

f=~/.config/muse/settings.json
mkdir -p ~/.config/muse
[ -f $f ] || echo '{"schema_version": 1}' > $f
jq '.endpoint_transport.base_url = "https://meta-api.int.exe.xyz/v1"' $f > $f.new && mv $f.new $f
echo 'export META_API_KEY=kept-by-exe-dev' >> ~/.bashrc
source ~/.bashrc
cd ~/project && muse exec "Reply with exactly: OK"

An environment key takes priority over a browser sign-in, so use one or the other. muse exec asks you to trust the folder the first time; run muse once there, or add --trust-workspace.

Note

We checked that Muse sends its requests to the integration's address with this setting, but haven't run it with a real Meta API key.

Troubleshooting#

• Read-only file system: the sandbox at work. Start muse from the folder it should write to; the rest of the disk, and .git inside the project, are read-only to sandboxed commands.
• failed to fetch model catalog: authentication failed: the key in META_API_KEY was rejected. It takes priority over a sign-in, so unset it to use your browser sign-in, or fix it.
• workspace trust ... is undecided from muse exec: run muse once in that folder and trust it, or pass --trust-workspace.
• A preview doesn't load: see "Peek at your app" above.
• invalid value 'unrestricted' for '--sandbox-network': Meta's docs use the new name; version 1.4.18 still wants enabled.
• Anything else: MUSE_TRANSPORT_TRACE=1 muse ... prints the requests to Meta's API, with credentials scrubbed, and Muse keeps its own logs in ~/.local/share/muse/local-tracing. /feedback reports a bug to Meta.