computer-use

Automates desktop GUI interactions via background input delivery and accessibility-tree element targeting.

Updated Jul 3, 2026
One-click install
npx skills add https://github.com/CHENHUI-X/toolbox --skill computer-use-chenhui-x
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: computer-use
Source: https://github.com/CHENHUI-X/toolbox/tree/main/official-skills/autonomous-ai-agents/computer-use
Command: npx skills add https://github.com/CHENHUI-X/toolbox --skill computer-use-chenhui-x

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Automating desktop applications usually requires moving the user's cursor and stealing keyboard focus, interrupting their work. This Skill drives the desktop in the background using accessibility-tree element indices instead of fragile pixel coordinates, so automation runs while the user keeps working. ## Core Features & Use Cases - Background GUI Automation: Capture screens with numbered element overlays (SOM mode), then click, type, scroll, and drag by element index without moving the user's cursor or raising windows. - Verify-and-Escalate Workflow: Every input action returns a structured verdict (confirmed, unverifiable, suspected_noop) with an escalation ladder from background element clicks to pixel coordinates to foreground delivery. - Cross-Platform Support: Works on macOS (AX), Windows (UIA), and Linux (AT-SPI) with any tool-capable model, powered by cua-driver under the hood. - Use Case: Ask the agent to fill out a form in a native desktop app like Figma or Mail while you continue typing in your editor; the agent captures the app, clicks fields by index, and verifies each step. ## Quick Start Use the computer_use tool to capture the Chrome window and click the Sign In button by its element index.

Frequently Asked Questions about computer-use

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I automate desktop app clicks without moving the user's cursor?

Use the computer_use capture action with mode="som" to get numbered element overlays, then call click with the element index. cua-driver delivers input in the background, so the OS cursor never moves and focus is not stolen.

What is the best way to click GUI elements reliably with an AI agent?

Click by accessibility element index rather than pixel coordinates. Capture first to get the AX-tree index, then pass element=N to the click action; indices are far more reliable across models than raw coordinates.

Does computer-use work on Windows, macOS, and Linux?

Yes, it supports all three platforms through cua-driver, mapping to each platform's accessibility framework: AX on macOS, UIA on Windows, and AT-SPI on Linux. Platform-specific deep dives ship in the cua-driver skill pack.

Why did my automated click have no effect on the desktop app?

Read the structured verdict returned by the action. If effect is suspected_noop or a refusal recommends escalation, retry by pixel coordinate, then with delivery_mode="foreground". Never silently repeat the same rung.

When should I not use computer-use for automation?

Avoid it for web page content automation (use headless browser_* tools instead), file edits (use read_file/write_file), and shell commands (use a terminal tool). Reserve it for native desktop apps and browser chrome.