cmux Cloud

Drive cmux Cloud machines (user-owned persistent Linux VMs) through the cmux CLI: create and attach, run commands and coding agents, move files, open ports and desktops, snapshot/fork/restore, and send notifications from inside a machine. Full per-command syntax lives in references/commands.md; cmux vm <sub> --help on the installed build is authoritative when they disagree.

Mental model

A machine is a persistent cloud VM owned by the signed-in cmux user. Its generated name (brave-otter) is its address everywhere; cmux vm rename sets a display label only. The cloud backend picks the provider; the machine outlives your panes, your Mac being closed, and reconnects.
Every machine runs a cmux session daemon (cmux-tui on current images, cmuxd-remote on older ones) that owns terminal sessions and scrollback. Clients (Mac panes, iOS) attach through short-lived leases minted by the backend; the transport depends on what the provider and image support. SSH is a fallback some providers and images cannot mint at all, and its absence is not an error.
New machines boot a desktop image (xfce + noVNC) plus a shell, with a persistent per-machine home. --base gives a shell-only machine.
Base is a separate single per-user persistent slot, pinned to the top of the sidebar. vm new mints fresh machines; vm base always reopens the same one.
Terminals on a machine live in its cmux-tui session (workspaces ws_…, terminals term_…). They keep running detached; cmux vm tree catalogs every surface (local and cloud) and every line is an address cmux vm open (machine targets) or cmux surface open (any entry, including This Mac) accepts, e.g. brave-otter/main/term_2f9c…, brave-otter:desktop, brave-otter:port/3000.
Pool machines (labeled agent-pool in vm ls) are provisioned by the vm run/vm agent router and reused for routed work. The router never drafts machines a person made by hand.
Plans cap active machine count and memory. cmux vm ls prints the meter (N of M machines on the <plan> plan) and, on free plans, when free cloud access expires.

Prerequisites

Host-side commands go through the running cmux macOS app's socket: the app must be running and signed in (machines belong to that account/team). cmux --version, cmux ping, then cmux vm ls is the fastest health check. Commands that document --json support it for scripting. Creating machines costs money and counts against the plan; prefer reuse over creation.

Workflows

Run a one-off command

cmux vm run -- uname -a                       # router picks/wakes/creates a pool machine
cmux vm run --sync -- bun test                # push current dir to work/<basename> first
cmux vm run --sync --pull work/app/dist -- sh -c 'cd work/app && bun run build'
cmux vm exec <id> -- pwd                      # when you already know the machine

exec gives up after ~35s client-side; vm run allows up to 15 minutes. For anything longer, start a detached terminal (vm agent, or cmux surface new-terminal --machine <id> -- <command...>) and check on it via vm tree / vm open.

Run a coding agent in the cloud

cmux vm agent --agent claude --sync -- "run the test suite and fix failures"
cmux vm agent --agent codex --machine brave-otter -- exec "summarize work/app"

Agents (claude, codex, opencode, pi are preinstalled) start as detached terminals in the machine's cmux-tui session: closing the pane does not kill them, and cmux vm open <machine>/<ws>/<term> reattaches from any device. Credentials for cloud agents come from cmux ai-accounts upload.

Create, attach, inspect

cmux vm new --detach                # fresh persistent machine; prints id + next commands
cmux vm shell <id>                  # terminal workspace on it (managed attach)
cmux vm desktop <id>                # its noVNC screen as a browser pane
cmux vm tui <id>                    # the machine's full cmux-tui client in a pane
cmux vm base                        # the persistent Base slot
cmux vm status <id>; cmux vm stats <id>; cmux vm ports <id>; cmux vm tools <id>

Preview a server running on a machine

cmux vm open <id> 3000              # private tokened URL, opened as a browser pane
cmux vm open <id> 3000 --print      # just print the URL

Move files

cmux vm push <id> ./myrepo work/myrepo      # dirs travel as tarballs, node_modules/.git excluded
cmux vm pull <id> work/report.pdf

Save and reproduce state

cmux vm snapshot <id> --name before-upgrade
cmux vm restore <snapshot-id>               # new machine from the snapshot
cmux vm fork <id>                           # copy of a running machine

Inside a machine

The guest cmux binary is the machine's relay CLI. Its commands are forwarded to the connected cmux app on the user's Mac, not executed in the VM: cmux notify posts to the user's notification center, cmux browser … drives the Mac's browser panes, cmux read-screen / send / workspace and pane verbs act on the Mac's UI. It resolves its socket from CMUX_SOCKET_PATH, then ~/.cmux/socket_addr, then the cloud CLI bridge socket; if no client is attached, relay commands fail — that is expected, not a broken install.

The one every cloud agent should know:

cmux notify --title "Build done" --subtitle "myrepo" --body "Tests green, artifact in work/dist"

--workspace/--surface default from CMUX_WORKSPACE_ID/CMUX_SURFACE_ID so the notification is attributed to the calling pane. Agent tool paths live under /root/.npm-global/bin, /root/.bun/bin, /root/.local/bin; use a login shell (bash -lc) when a tool is "missing".

Rules

MUST NOT run cmux vm rm on a machine you did not create in this session unless the user names it this turn. It is irreversible and unprompted. When the user wants a fresh Base, use cmux vm base reset, which retains the old VM.
MUST NOT pass a name to cmux vm new — it takes no positional arguments (rejected to prevent accidental paid provisioning). Create, then cmux vm rename.
Prefer vm run/vm agent routing over creating machines; creation counts against the plan limit and bills the user. Check cmux vm ls before vm new.
Do not loop on vm status waiting for readiness; use cmux vm wait <id> [--wake].
Do not use cmux vm ssh unless the managed attach path failed; some providers and images cannot mint SSH, so treat "SSH unavailable" as normal, not as an error to fix. Say why in the handoff when you fall back.
Do not call the internal attach helpers (vm ssh-attach, vm-pty-attach, vm-ssh-attach, vm-pty-connect, vm-tui-connect, vm-tui-approve).
vm exec quoting is faithful per argv element; do not pre-quote. Wrap shell constructs as -- sh -c '<script>'.
Provider choice belongs to the backend. Use --provider only for a rollback or experiment the user asked for.
Use --json only on commands whose --help documents it; interactive verbs (shell, tui, desktop) have none. Do not parse the human table output either way.
Plan limits, sizes, and providers change; read them from cmux vm ls and --help output rather than assuming values from this file.

cmux-cloud