A Claude Code plugin carrying the operating knowledge for the built-in computer-use MCP server, the desktop screen-control surface. Two skills, one concern: making a screen-control session succeed, and making its failures legible.
| Skill | What it does |
|---|---|
/computer-use:diagnose |
Symptom-to-cause for a screen-control session. Covers the screenshot pixel budget and when zoom recovers detail, the ladders for capture and input failures, and the per-OS quirks that make synthesized input behave unlike a human's. A symptom guide answers the common cases with no file load. |
/computer-use:setup |
Read-only preflight. Confirms the surface and tool availability, then reports the screensaver, display-sleep, and lock timeouts that end a session mid-run regardless of how busy Claude is. Check-only: every prerequisite is external or a system setting this plugin must not write. |
The computer-use tool descriptions already document the mechanics: the actions, the allowlist gate, the permission tiers, the ladder that prefers a dedicated MCP over Chrome over screen control. This plugin restates none of that, on purpose: a standing instruction that repeats what the surface already says is a per-session tax for zero gain.
What it carries instead is the material that is not derivable from the tool surface and is not documented upstream:
- The idle timer does not see Claude. Synthesized input does not reset the OS idle timer, so a long session hits the screensaver or display-sleep timeout no matter how much Claude is clicking, and cannot recover itself once it does. Measured, not inferred.
- Screenshot quality is a fixed pixel budget, not a scale factor, so the intuitive fixes
(smaller monitor, lower resolution) do not work and one non-obvious one (
zoom, which re-captures at full resolution) does. - Capture and input failures are environment state, not tool defects, and each has a distinct signature worth reading rather than retrying past.
- No machine facts baked in. Focus-stealing utilities, monitor layouts, and timeout values are described as classes with general mitigations; specific processes and displays belong in the operator's own memory file, not here.
- Per-OS content is opt-in. The skill routes to a quirks file for the platform you are on and loads nothing else.
- Zero configuration. No
userConfig, no tracked project files, nothing to write.
Every empirical claim in this plugin was measured on Windows 11 with the Claude Desktop surface on 2026-08-10. macOS is supported by the platform, and the platform-neutral content applies there, but no macOS quirks file ships because none has been verified. Rather than generate plausible macOS specifics, the plugin states the gap and serves macOS users the platform-neutral material only.
Surfaces are not interchangeable, and the skill says so: computer use in the Claude Code CLI is macOS-only, while the Desktop app covers macOS and Windows. Guidance written for one is wrong on the other.
- No system-setting changes. Power, screensaver, and lock policy are reported with their measured values and handed to the operator. The plugin never writes them.
- No
applyaction. There is nothing it could conformingly write. The check-only carve-out applies, sosetupverifies and reports and stops. - No mechanism claims it cannot support. Where a behavior is reproducible but unexplained, such as Windows shell menus ignoring a synthesized Escape, the plugin says exactly that and ships the rule anyway.
The computer-use MCP server must be enabled, on a plan that supports it (research preview; Pro
or Max, not Team or Enterprise), with OS-level permissions granted. /computer-use:setup probes
all of it and reports what is missing; installing and enabling remain the operator's.