Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add thewolffish/wolffish-app --skill computer-usegit clone --depth 1 https://github.com/thewolffish/wolffish-appWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/thewolffish/wolffish-app/computer-use)<a href="https://agentmods.dev/skills/thewolffish/wolffish-app/computer-use"><img src="https://agentmods.dev/badge/skills/thewolffish/wolffish-app/computer-use/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/thewolffish/wolffish-app/computer-use"><img src="https://agentmods.dev/badge/skills/thewolffish/wolffish-app/computer-use.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00066 | $0.11407 |
| Opus 5.5 | $0.00026 | $0.04563 |
| Sonnet 5.5 | $0.00013 | $0.02281 |
| Haiku 4.5 | $0.00007 | $0.01141 |
Grade C, and why
computer-use scanned grade C with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Recursive force deletehighDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
- pattern: 'computer_(keyboard_type|set_value).*(sudo|rm -rf|password|secret|token)' How it starts
The opening of the file, as written. The whole thing — 908 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Computer Use — Verified Aim, Background Delivery, Two Grounding Routes
The agent sees and controls the desktop through a closed feedback loop designed for surgical accuracy with any vision-capable model — and, since v3, without taking the user's mouse away from them.
What the model works with
- One coordinate space, owned by the plugin. Every image a tool returns (display screenshot, window screenshot, zoom, or magnifier) becomes the current frame. The model gives coordinates as pixels read off the latest image; the plugin does all translation — downscaling, Retina/HiDPI, multi-monitor offsets, and the window-local capture pixels the native driver expects. The model never does coordinate math.
- Two grounding routes. Pixels (screenshot → zoom → aim → click) and elements
(
computer_findby name through the app's accessibility tree →computer_click_elementby token). On every pixel click the element the app reports under the point is echoed back, so a wrong aim is caught by one accessibility query instead of by eyesight. - Background delivery through a native driver. Clicks, keys, scrolls and drags are
posted to the target window (cua's Rust driver,
@trycua/cua-driver, MIT): the user's pointer does not move and the window is not raised. When an app refuses background input the driver answers with a code (background_unavailable,background_occluded,background_uipi_blocked) and the plugin escalates one rung — foreground delivery with the pointer restored — and, only if that is impossible, the legacy path that moves the real pointer (nut-js). Every result names the rung that ran. - Evidence on every action. Delivery rung, the driver's effect verdict
(
confirmed,partial,unverifiable,suspected noop,refused), whether the real pointer moved during a foreground action (the user is using the mouse), the element under the point, and an objective before/after pixel-change measurement. The action's stated intent (target,expect) rides the runtime tail into the next step, so the model verifies before it plans — one forward pass, no extra call. - A shadow cursor in the screen indicator: it glides to the point before the input
is posted and pulses on the press, so the person sees cause before effect while their
own pointer stays free.
computer_mouse_moveaims the shadow cursor only;computer_hoveris the one tool that moves the real pointer (tooltips, hover menus).
What ships with it
6 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 1mo ago Changed · -6 lines · -27 tokens per session 720efd8108c7
- 1mo ago Changed · +21 lines baba68502d68
- 1mo ago Changed · +388 lines · +40 tokens per session b8fc31fa2040
- 1mo ago Changed · +27 lines c5929df92e82
- 1mo ago Changed · +3 lines 00ee42a2fa70
- 1mo ago Changed · +73 lines 1f5f907cd170
- 1mo ago First seen · 402 lines · 53 tokens per session scan C efd9392c5fd6
computer-use is a skill published in the GitHub repository thewolffish/wolffish-app (5 stars, last pushed 22d ago), licensed MIT. It adds 66 tokens to every session and 11,407 once invoked, about $0.0003 per session on Opus 5.5. A static security scan graded it C with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
agent-computer-use
REQUIRED for any task that involves operating a desktop application — opening apps, clicking buttons, typing into fields, pressing keys, scrolling, dragging, reading what's on screen, moving or resizing windows, or verifying state after an action. Always use the agent-cu CLI commands (open, snapshot, click, type, key…
claude-api
Build, debug, and optimize Claude API / Anthropic SDK apps. Apps built with this skill should include prompt caching. Also handles migrating existing Claude API code between Claude model versions (4.5 → 4.6, 4.6 → 4.7, retired-model replacements). TRIGGER when: code imports anthropic/@anthropic-ai/sdk; user asks for…
create-skill
Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.
figma-generate-design
Use this skill alongside figma-use when the task involves translating an application page, view, or multi-section layout into Figma. Triggers: 'write to Figma', 'create in Figma from code', 'push page to Figma', 'take this app/page and build it in Figma', 'create a screen', 'build a landing page in Figma', 'update the…
figma-create-design-system-rules
Generates custom design system rules for the user's codebase. Use when user says "create design system rules", "generate rules for my project", "set up design rules", "customize design system guidelines", or wants to establish project-specific conventions for Figma-to-code workflows. Requires Figma MCP server…
figma-generate-library
Build or update a professional-grade design system in Figma from a codebase. Use when the user wants to create variables/tokens, build component libraries, set up theming (light/dark modes), document foundations, or reconcile gaps between code and Figma. This skill teaches WHAT to build and in WHAT ORDER — it…