What does this skill do, and when should you use it?
cmux-cloud-vm is one of 25 skills bundled in the cmux repository and teaches an agent how to manage cmux Cloud virtual machines. It prescribes a workflow of discovering machines, authenticating, selecting a target, and executing work either as bounded commands or detached agents, with headless operation and reuse of existing machines and workspaces emphasized throughout. Machine terminals persist when panes close or the Mac disconnects, making them suitable for long-running agent work. The backend implementation belongs to cmux-backend; this skill covers client-side operation only.
- Runs
cmux vm --help,cmux auth status,cmux vm ls --, andcmux vm route --to discover machines and placement;routeinspects placement without creating a machine - Executes bounded commands with
cmux vm run --machine <id> -- <command>and retains the exit code - Launches detached coding agents with
cmux vm agent --machine <id> --agent codex --no-open -- "<task>" - Observes headlessly via
cmux vm tree,terminal read,terminal wait-exit, andterminal outputfor completion and full output - Presents verified results with
vm open/workspace open; noteworkspace rmkills terminals whileworkspace closeonly detaches them - Consults reference docs (commands.md, guest.md, agent-workflows.md) for exact flags, JSON/exit contracts, limits, and guest operations
- cmux users running parallel Claude Code/Codex sessions who need to offload long-running tasks to a persistent cloud terminal without occupying their Mac
- Agents that must run builds or tests longer than a session, where the machine terminal should survive disconnects
- Developers bootstrapping a project dev environment on a cloud machine, reviewing `cmux vm dev <id> --dry-run --` before executing the plan
- Scenarios where a finished task's workspace should be surfaced to the user as a pane or shared URL for inspection
- Multi-machine users who want to prefer the router's pool machine or pin an existing target instead of provisioning new ones
- Non-macOS users: host operations require the cmux app, sign-in, and its private tunnel, and cmux is macOS-only
- Users without a cmux Cloud account or access: the skill does not provision cloud resources itself
- Anyone looking for the cmux backend implementation: SKILL.md states backend code belongs to cmux-backend; this is a client-side operating guide
How do you install this skill?
- The skill only works with the macOS cmux app, a signed-in account and the WireGuard tunnel up, and depends on the cmux cloud backend; mainland-China users should verify network reachability first.
- License metadata is NOASSERTION and the publisher is unverified; plan-tier restrictions (free plan requires Pro to create machines) should be confirmed before adoption.
- Destructive commands (vm rm, base reset, workspace rm, domains rm) are permanent; agents must strictly follow the documented authorization and confirmation requirements.
- The commands.md reference is truncated in the provided material; check the full file and the installed CLI's --help before use.
- Shell / CLI
- Network access
- Local filesystem
cmux CLIcmux macOS app (for host operations)cmux Cloud account/sign-in
The source documents no dedicated skill-install commands. The skill lives at skills/cmux-cloud-vm/SKILL.md in the manaflow-ai/cmux repository; copy the folder into your client's skills directory per the Agent Skills convention. Claude Code (no specific command given in the source; generic skills-directory approach)
git clone https://github.com/manaflow-ai/cmux.git
cp -r cmux/skills/cmux-cloud-vm ~/.claude/skills/Prerequisite for host operations: the cmux macOS app, installed and signed in.
brew tap manaflow-ai/cmux
brew install --cask cmuxHow do you use this skill?
Once installed, send your agent any of these to trigger it:
- List my available cmux Cloud machines and tell me which pool machine I can reuse
- Start a detached codex agent on machine <id> using cmux vm to refactor the login page
- Dry-run the dev layout for this project on a cloud machine and show me the JSON plan before executing
- Wait for the terminal on machine <id> to exit, then show me the full output and open its workspace
Triggers when a user asks for cmux vm/cloud tasks. Flow: start with discovery (cmux vm ls --, cmux vm route --), reuse the router's pool machine or pin a target with --machine <id>, preferring another workspace on an existing machine over a new machine; then choose an operation (bounded command, detached agent, dev layout, headless observation); finally present results with vm open/workspace open. While working, use --no-open, --detach, or --print to avoid stealing focus; a finished wait-exit is not proof tests passed — read the actual output.
cmux vm route --
cmux vm run --machine <id> -- <command>
cmux vm terminal wait-exit <id> <term>Constraints: destructive actions (deletion, reset, resize) need user authorization for the target and effect; keep account tokens on the host and never copy credentials into a machine unless requested.
What are this skill's strengths and limitations?
- Machine terminals persist when panes close or the Mac disconnects — well suited to long-running agent work
- Encourages reuse of existing machines and the router pool, avoiding unnecessary provisioning and deletion
- Explicit operational constraints on authorization, credential safety (no copying tokens into machines), and exit-code verification
- Ships reference docs (commands, guest, agent-workflows) covering exact flags and JSON contracts
- Tightly coupled to the cmux ecosystem: requires the cmux app, sign-in, private tunnel, and a Cloud account; not standalone
- macOS host only; other platforms are unverified in the source
- Repository license is NOASSERTION: the core is GPL-3.0-or-later, but the server/Cloudflare-worker components use Business Source License 1.1, and production use requires a commercial license
- No test suite or dedicated installation documentation for the skill itself in the source
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| cmux Cloud VM Skill this page | 56 · Use with care | ★ 28k | 1d ago | NOASSERTION |
| Cua Spaces (cua-spaces skill) | 46 · Use with care | ★ 29k | 1d ago | MIT |
| cmux Computer Use Skill | 50 · Use with care | ★ 28k | 1d ago | NOASSERTION |
| Cua Driver GUI Automation Skill | 61 · Recommended | ★ 29k | 1d ago | MIT |
| Storage Analyzer | 60 · Recommended | ★ 21k | 10d ago | MIT |
Compared to tmux: tmux is a terminal multiplexer running inside any terminal; cmux is a native macOS app with built-in vertical tabs, split panes, an embedded browser, and a socket API, with no config files or prefix keys. cmux can also attach to remote tmux sessions over SSH, or use a persistent Cloud VM when work must stay online while the Mac is offline.
How did FollowSkills review this skill?
SKILL.md and references explicitly require user authorization for destructive operations (vm rm, base reset, workspace rm), keep credentials on the host, forbid secrets in layouts/logs/command args, default domain publications to personal and warn that public URLs are bearer credentials, use SHA-256 checks and pinned host keys for push, and support network allowlists. Deductions: static review cannot verify these claims match implementation; publisher unverified, license metadata NOASSERTION, attribution and external backend effects not fully auditable.
Very high internal consistency: verb catalog, exit-code contracts, -- payloads, and failure feedback (wait-exit codes, 'a finished wait is not proof tests passed') are all specified, with guidance to read --help first. Deductions: static cap of 10; the cited tests/test_cloud_vm_skill_coverage.py is not in the provided material; key paths (routing, provisioning, tunnel) not executed; some examples assume netstat/wget availability.
Description gives clear triggers (cmux vm/cloud), non-fit boundaries are declared (backend belongs to cmux-backend), guest/host grammar is distinguished, and a read-only preflight guides discovery. Deductions: hard dependencies on the macOS app, sign-in, WireGuard tunnel and the cmux cloud backend; complex plan-tier limits; core function entirely depends on overseas cloud services with questionable mainland-China reachability and no declaration; no Chinese support.
Good layered information architecture: SKILL.md as entry, task-indexed references, anchored command catalog, high-quality guest grammar table and 1:1 sidebar parity table, with known limitations disclosed (legacy flags, deprecated peer enrollment). Deductions: license metadata NOASSERTION; no version/changelog binding (only 'current source revision'); ownership and update path unclear; commands.md is truncated in the material.
Documentation is actionable end-to-end (route → run/agent → wait-exit → output → open workflow, idempotent dev-server script, fork-based parallel experiments) with clear marginal value over manual SSH/tmux. Deductions: static cap of 7; no verified actual outputs; limited evidence on cost/benefit given paid-plan dependencies; sample jq fields unverified by execution.
Claims trace to in-repo commands and socket method names; the parity table marks verification status and a coverage test file is referenced — auditable primary material. Deductions: static cap of 5; no third-party execution evidence; the 'Verified ✅' column is author-asserted; no cross-source corroboration; fact/inference boundaries need human re-check.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →