What does this skill do, and when should you use it?
cmux Core Control is a skill file that teaches an AI agent to drive terminal topology through the cmux CLI. cmux is an open-source Ghostty-based macOS terminal with vertical tabs, split panes, and notifications for AI coding agents. The skill defines a layered handle model — window, workspace, pane, and surface (a tab within a pane) — and provides concrete commands for creating workspaces and splits, moving and reordering surfaces, running workspace context-menu actions, and flashing an attention cue. It is aimed at automation that needs deterministic placement and navigation in a multi-pane layout. It is a documentation-only skill: no scripts, all actions are issued as shell commands.
- Runs
cmux identify --to self-identify the caller's context in the window/workspace/pane/surface hierarchy. - Lists and creates topology elements:
list-windows,list-workspaces,list-panes,list-pane-surfaces,new-workspace,new-split,new-pane,new-surface. - Moves and reorders surfaces with
move-surface,split-off, andreorder-surface, routing focus with--focus. - Launches initial commands on new terminals via
--command <text>(the shell stays alive after the command exits). - Customizes workspaces with
workspace-action --action set-color|set-description|...and flashes an attention cue withtrigger-flash --surface surface:N. - Reloads
~/.config/cmux/cmux.and Ghostty config withcmux reload-config, and inspects browser download history withcmux browser ... download list.
- A developer running parallel Claude Code or Codex sessions who wants an agent to place each task into its own workspace or split automatically.
- An automation engineer scripting layouts — e.g. open the dev server in a right pane and an agent session below — before handing control to a human.
- An agent integration that needs to signal a human without interrupting: trigger a flash cue on the specific surface doing the work.
- A user reorganizing an existing cmux layout from the CLI: reorder or move surfaces, rename workspaces, set colors and descriptions.
- A team that wants the agent to verify its own caller context (which pane/surface am I in) before issuing layout commands on macOS.
- Non-macOS users: cmux is a macOS-only native Swift/AppKit app with no other platform build.
- Anyone not running the cmux app: every command targets a live cmux instance — Ghostty, tmux, or zellij alone do not provide it.
- Callers without shell execution: the skill is entirely CLI-based and does nothing in a chat-only client with no shell access.
How do you install this skill?
- The skill requires the cmux app and cmux CLI on macOS; it is unusable without them — verify your environment first.
- --command injects literal text into a new shell where the shell interprets it; combining it with untrusted input creates an injection surface — pass controlled text only.
- Static review executed nothing; documented CLI behavior must be verified by the user in practice.
- Skill metadata license is NOASSERTION while the repo is GPL-3.0 (web/ is BUSL-1.1); check commercial compliance yourself.
- Always make the recommended timestamped .bak of cmux. before edits, or recovery from a broken config may not be possible.
- Shell / CLI
- Local filesystem
cmux macOS app (installed and running)
The skill lives at skills/cmux/SKILL.md in the manaflow-ai/cmux monorepo (the repo bundles 25 skills). Install the skill folder the way your host normally does; no skill-specific install command is documented in the source. Claude Code — clone the repo and place the skill folder in your host's skills directory:
git clone https://github.com/manaflow-ai/cmux
# copy skills/cmux/ into your host's skills directoryPrerequisite: install the cmux app itself:
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:
- Use cmux to create a new workspace, run `npm run dev` in a right split, and focus it
- List my current windows, workspaces, and panes, then tell me which surface I'm on
- Move surface:7 into pane:2 with focus and trigger a flash on it
- Set my current workspace's description to 'Ship checklist' and its color to Blue using cmux
The host agent triggers the skill when automation needs deterministic placement or navigation in a cmux layout. The flow: discover the caller's context with cmux identify --, enumerate topology with the list-* commands, then mutate it with new-workspace / new-split / move-surface / reorder-surface / workspace-action / trigger-flash. Output defaults to short refs (window:N, workspace:N, pane:N, surface:N); pass --id-format uuids|both for UUIDs. Notable options: --command is accepted on terminal surfaces only (rejected with --type browser|simulator|agent-session, ignored by new-workspace --layout); cmux reload-config refreshes both cmux. and Ghostty config without an app restart. Deep-dive docs live in the references/ folder alongside the skill.
What are this skill's strengths and limitations?
- Precise, deterministic command set covering layout creation, focus routing, reordering, and attention cues.
- Clear handle model (window/workspace/pane/surface) with short refs and optional UUID output.
- Well documented via references/ docs and cross-links to sibling skills (settings, browser, workspace, capture).
- Includes safety guidance, e.g. backing up cmux. before edits and the boundary with Ghostty-owned rendering settings.
- Strictly macOS + cmux app; the skill is useless without both.
- Documentation-only — no scripts or tests, so command behavior cannot be verified against a specific cmux version.
- The repository's license field is listed as NOASSERTION (the codebase is GPL-3.0-or-later, but server components fall under BUSL 1.1), so license classification needs checking.
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 Core Control this page | 47 · Use with care | ★ 28k | 1d ago | NOASSERTION |
| cmux Workspace Skill | 64 · Recommended | ★ 28k | 1d ago | NOASSERTION |
| cmux Custom Sidebar | 50 · Use with care | ★ 28k | 1d ago | NOASSERTION |
| cmux Settings Management Skill (cmux-settings) | 58 · Recommended | ★ 28k | 1d ago | NOASSERTION |
| cmux Customization | 55 · Use with care | ★ 28k | 1d ago | NOASSERTION |
cmux explicitly positions itself against tmux: tmux is a terminal multiplexer running inside any terminal, while cmux is a native macOS GUI app with vertical tabs, splits, an embedded browser, and a socket API. This skill drives cmux's CLI, not tmux's prefix-key interface.
How did FollowSkills review this skill?
The skill only manipulates local terminal topology and config via the cmux CLI: no network exfiltration, no credential reads. Docs explicitly instruct copying cmux. to a timestamped .bak before edits and disclose legacy fallback config paths, giving reasonable data-flow transparency. Deductions: backup is advice, not enforced; --command injects arbitrary text into a new shell; permission/confirmation scoping is not addressed at the skill level; publisher identity is unverified.
SKILL.md and references are self-consistent on --command semantics (literal delivery, blank ignored, type-conflict errors, remote tmux rejection), and the handle model and command examples align. Deductions: static review cannot execute; failure-output formats and error feedback under abnormal input are not shown; edge paths (identify resolution failure, post-health-check behavior) are uncovered.
Scenario is clear (deterministic placement/navigation in multi-pane layouts), with partial boundary declarations (macOS-only, --command type limits, remote tmux rejection) and a precise trigger description. Deductions: the prerequisite that the cmux app and CLI are on PATH is not stated in the skill; no Chinese-language support statement; mainland-China reachability is undiscussed (minor since the skill is local).
Good layering: quick-start SKILL.md plus five topic references plus sibling-skill cross-links, a 'When to Use' table showing progressive disclosure; repo has GPL-3.0 LICENSE, SECURITY.md, and real CI workflows. Deductions: skill metadata license is NOASSERTION; the skill itself has no version or changelog; skill-level maintenance responsibility and update path are undeclared, and shared repo evidence cannot be attributed directly to this skill.
The claimed output (deterministic control of windows/workspaces/panes/surfaces via CLI) maps one-to-one to documented commands, with an openai.yaml interface descriptor and clear marginal value for agent automation. Deductions: static review cannot confirm the commands actually run; no representative output samples or comparative demos; value presupposes the cmux app is installed.
Auditable primary material exists (SKILL.md, references, openai.yaml, LICENSE, SECURITY.md, CI workflows) with cross-document corroboration. Deductions: nothing was executed; the repo's CI and tests cover the app and auth packages, not this skill's key paths; no independent third-party reproduction, so the 5-point static cap applies.
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 →