What does this skill do, and when should you use it?
cmux-capture is a skill bundled in the manaflow-ai/cmux repository that teaches an agent to use `cmux shot` and `cmux record` to screenshot or record cmux's own windows. Output formats include png/jpeg/mp4/gif, and both commands work in release builds and in CI, going through a permission-free path that never requests macOS Screen Recording access. Captures only include cmux windows, so another app sitting in front does not appear in the frames. The skill targets changes that are visual and hard to describe in text — drags, animations, layout shifts — where a reviewer needs evidence rather than a claim.
- Runs
cmux shotto write a png/jpeg of a cmux window, printing width x height, bytes, and path - Runs
cmux recordto write an mp4/gif clip that auto-stops at --max-seconds (default 15, max 120) - Draws captions into the clip via
cmux record notefrom that moment on - Reuses the same --region x,y,w,h rectangle across shot and record to go from a clip back to a still
- Returns full responses with --, including state, frame counts, and fps_effective
- Works with scripts/pr-media.py to upload clips or convert an mp4 to a gif for a PR
- A developer changes a UI layout and has the agent attach a still screenshot to the PR as proof
- An agent performs a sidebar drag or animation and records a gif so the reviewer sees the motion
- A CI dogfood scenario wraps key steps in a record step with note captions for reviewable evidence
- A reviewer doubts a textual claim, so the same region used in a clip is shot as a final-state still
- An agent recording itself sets max-seconds so the clip never keeps running after it stops paying attention
- Non-macOS users: cmux is a macOS-only native app and the skill depends on the cmux CLI
- Capturing anything outside cmux: both commands capture only cmux's own windows
- General-purpose screen capture of arbitrary apps or the whole screen: use built-in OS tools instead
How do you install this skill?
- This is a static source review; no commands were executed and confidence is low.
- The skill works only on macOS with the cmux app installed; it is entirely unusable without cmux.
- Downloading cmux relies on GitHub Releases, which may be unreliable from mainland-China networks.
- Repository license metadata is NOASSERTION, inconsistent with the GPL-3.0-or-later LICENSE file; verify licensing independently before use.
- The publisher is unverified by the FollowSkills curated registry; identity is unknown.
- Documented behavior has not been verified against the implementation; defer to cmux --help and observed behavior.
- Shell / CLI
- Local filesystem
cmux macOS app with cmux CLI (macOS only)
The skill lives at skills/cmux-capture/SKILL.md in the manaflow-ai/cmux repo, one of 25 skills bundled there. No dedicated skill-install command is documented in the source; the prerequisite is installing the cmux app itself:
macOS (DMG / Homebrew, installs the cmux app)
brew tap manaflow-ai/cmux
brew install --cask cmuxStandard Agent Skills practice is to copy the skills/cmux-capture folder into your agent client's skills directory; the exact path depends on your client and is not specified in the source.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- I just changed the settings sheet layout — take a png of the cmux window and put it in the PR description
- Record an 8-second gif of the sidebar drag and add a caption right before the drag starts
- Use --region 0,0,420,900 to record the sidebar for 6 seconds, then shoot the same region as a final-state still
- Check whether a recording is running, and if not, start one of the workspace-switching animation
The skill is triggered by an agent when the task is visual, when a drag or animation cannot be described in text, or when a reviewer needs evidence rather than a claim. Core flow:
cmux shot --label settings-sheet
cmux shot --jpg --max-width 1200 --out ~/before.jpg
cmux record start --gif --max-seconds 8 --label sidebar-drag
cmux record note "dragging the workspace"
cmux record stopKey options: screenshot for still things (sheets, menus, layouts), record for changes over time (drags, animations, reorders); only one recording at a time — a second start returns conflict; --gif already picks 8 fps and half scale; --out extension must match the format; frames are timed by capture time, so check fps_effective in -- if a clip looks slow. In CI, use the record step table in cmux-testing/references/dogfood-scenarios.md. Full flag reference: references/commands.md.
What are this skill's strengths and limitations?
- Never requests macOS Screen Recording permission — a completely permission-free path
- Works in release builds and CI, making it suitable for automated verification
- Captures only cmux windows, so overlapping apps do not pollute the frames
- Recordings self-stop at --max-seconds, so clips are never left running
- Shot and record share the same region coordinate system, letting a clip detail be re-shot as a still
- Hard dependency on the cmux app and macOS — useless on other platforms or terminals
- Cannot capture anything outside cmux's own windows
- Repo license is listed as NOASSERTION (GPL-3.0-or-later core with BUSL 1.1 server components) — verify scope before commercial use
- Effectiveness depends on having cmux installed locally; the source provides no cross-environment test evidence
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-capture: Screenshot and Record cmux Windows this page | 55 · Use with care | ★ 28k | 1d ago | NOASSERTION |
| cmux Shared Behavior Rules | 48 · Use with care | ★ 28k | 1d ago | NOASSERTION |
| cmux Workspace Skill | 64 · Recommended | ★ 28k | 1d ago | NOASSERTION |
| cmux Diagnostics | 51 · Use with care | ★ 28k | 1d ago | NOASSERTION |
| cmux Core Control | 47 · Use with care | ★ 28k | 1d ago | NOASSERTION |
The cmux FAQ contrasts cmux with tmux: tmux is a multiplexer inside any terminal, cmux is a native macOS GUI app. For this skill specifically, the natural alternatives are macOS built-in screenshotting (screencapture) and generic screen recorders — but those require Screen Recording permission and capture the whole screen. cmux-capture's differentiators are being permission-free, capturing only cmux windows, and working in CI.
How did FollowSkills review this skill?
The skill only screenshots/records cmux's own windows, explicitly states no Screen Recording permission is involved or requested, discloses output-path handling (hidden sibling, atomic move, directory refusal), and shows good least-privilege and data-flow transparency. Deductions: no verified publisher, repository license metadata is NOASSERTION (though LICENSE file is GPL-3.0-or-later), and no documented rollback for files already produced.
The docs are highly self-consistent: flag tables, error codes, limits, and edge cases (region outside window, gif frame ceiling, extension mismatch) are explicit, with diagnosable, intelligible failure modes. But this is a static review only — no commands were executed, no happy-path reproduction, and no test evidence covering this CLI's key paths was found, so it cannot exceed 10.
The description gives clear semantic triggers (visual change, drag/animation, reviewer needs evidence), and the body offers a shot-vs-record guide, agent notes, and CI usage. Deductions: it hard-depends on macOS and an installed cmux app, and its non-fit boundary is not declared inside the skill; downloading the app from GitHub Releases may be unreachable from mainland-China networks, with no caution provided.
Good layered architecture: main SKILL.md progressively discloses into references/commands.md covering every flag, response field, and error code, plus a practical selection guide. Deductions: no skill version or changelog; repository license metadata (NOASSERTION) contradicts the actual GPL-3.0-or-later LICENSE; maintenance ownership and update path are only implied by repository activity, not stated in the skill.
The skill directly enables the screenshot/recording task with directly usable example commands, fully defined output formats (-- fields, error codes), and clear added value over manual capture (region reuse, captions, CI dogfood integration). But static review cannot verify actual output usability, and marginal value depends on the user already having cmux installed, so it stays below the 7 anchor.
Key claims are cross-checkable against cmux shot --help, and the repository has real CI workflows and test infrastructure, but no test evidence covering the capture key paths was found, and no execution was performed. This stays at the static cap of 5 minus a deduction for the absence of third-party execution evidence specific to this skill.
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 →