convez.ai
Agente
zgyanye

windows-computer-use

zgyanye/windows-computer-use

Ver no GitHub
InicialPluginScore · de 0 a 100

Windows Desktop Agent for Claude Code

Instalação

$npx convez-ai zgyanye/windows-computer-use

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

windows-computer-use

Windows Desktop Agent for Claude Code — perceive, act, and transfer on Windows.

Zero npm dependencies. Pure Node.js + PowerShell. Works with any Chromium browser (Chrome, Edge, QQ Browser, 360, Brave, etc.).

English | 中文


Why

Anthropic's official Computer Use is macOS only. Windows users have no first-party solution.

This skill fills that gap — and goes further:

  • CDP-based DOM reading (0.5s, structured) instead of screenshot+OCR (3-5s, lossy)
  • Window-level screenshots (~200KB) instead of full-screen (~2MB)
  • Full input control — mouse, keyboard, scroll, drag
  • One command install, one command setup

Quick Start

Prerequisites

  • Windows 10 or later
  • Node.js 21 or later (node --version)
  • A Chromium browser (Chrome, Edge, QQ Browser, etc.)

Install

# Clone into your Claude Code skills directory
git clone https://github.com/YOUR_USERNAME/windows-computer-use.git "%USERPROFILE%\.claude\skills\windows-computer-use"

Or via skills.sh (coming soon):

npx skills add YOUR_USERNAME/windows-computer-use

One-Time Setup

node scripts/config.cjs --setup

This patches your browser shortcuts to enable CDP debugging. After this, your browser supports direct DOM reading with zero overhead.

Usage

In Claude Code, just talk normally:

You sayWhat happens
"看页面" / "what's on screen"Reads current browser page DOM (0.5s)
"微信说了什么" / "read WeChat"Captures WeChat window, analyzes with vision
"这是什么" / "identify this"Auto-screenshot + vision identification
"切到微信" / "switch to Code"Activates target window
"点击那个按钮" / "click the button"Mouse click at coordinates
"输入 XXX" / "type hello"Keyboard input
"复制到剪贴板" / "copy to clipboard"Writes to clipboard

Architecture

screen.cjs (perceive)
├── CDP (Chromium DevTools Protocol) → 0.5s DOM read
├── UIA (Windows UI Automation) → app window enumeration
└── Screenshot (window-level, not full screen) → vision analysis

control.cjs (act)
├── switch → activate windows by title/process
├── click/dblclick/rightclick → mouse at coordinates
├── drag → click-and-drag
├── scroll → mouse wheel
├── type → keyboard text input
└── key → special keys (enter, ctrl+c, etc.)

clipboard.cjs (transfer)
├── read → auto-detect text/image/files
└── write → text, image file, or file path

config.cjs (setup)
├── check → CDP status, browser shortcuts
└── setup → auto-patch shortcuts for CDP

Comparison

Anthropic Computer UseThis Skill
PlatformmacOS onlyWindows
Browser readingScreenshot + OCRCDP (DOM, 0.5s)
ScreenshotFull screenTarget window only
InputMouse + keyboardMouse + keyboard + scroll + drag
DependenciesManaged by AnthropicZero npm deps
LicenseProprietaryMIT

Supported Browsers

All Chromium-based browsers:

  • Google Chrome
  • Microsoft Edge
  • QQ Browser (QQ浏览器)
  • 360 Browser
  • Brave
  • Opera
  • Any Chromium fork

Firefox support via WebDriver BiDi planned.

Troubleshooting

SymptomFix
CDP not connectingRun node scripts/config.cjs --setup
Browser not detectedRestart browser via the modified shortcut
Screenshots blankCheck DPI scaling; try --mode=uia first
Permission errorsRun terminal as regular user (not admin)
Chinese paths garbledSet terminal encoding to UTF-8: chcp 65001

Security

  • All CDP communication is localhost-only (127.0.0.1:9222)
  • No network exposure, no cloud dependency
  • Input simulation uses standard Win32 APIs (SendInput, SendKeys)
  • No kernel drivers, no DLL injection
  • Source code is fully auditable (~600 lines total)

Contributing

Issues and PRs welcome. Key areas for contribution:

  • Firefox/Gecko WebDriver support
  • Multi-monitor coordinate handling
  • Non-Chinese Windows locale testing
  • Accessibility improvements (MSAA fallback)

License

MIT — do whatever you want, no warranty.


中文说明 {#chinese}

这是什么

Windows 桌面智能代理 Skill — 让 Claude Code 能看你的屏幕、操作你的电脑、读写你的剪贴板。Anthropic 官方的 Computer Use 只支持 macOS,这个填补了 Windows 的空白。

核心能力

能力方式速度
读浏览器内容CDP 直读 DOM0.5秒
识别屏幕内容窗口截图+vision2-4秒
切换窗口Win32 API<1秒
鼠标操作SendInput即时
键盘输入SendKeys即时
剪贴板读/写 文字+图片+文件<0.5秒

安装

git clone https://github.com/YOUR_USERNAME/windows-computer-use.git "%USERPROFILE%\.claude\skills\windows-computer-use"

配置

node scripts/config.cjs --setup

会自动修改你桌面的浏览器快捷方式,加上 CDP 调试端口。一次配置,永久生效。

要求

  • Windows 10 或更高
  • Node.js 21 或更高
  • Chrome/Edge/QQ浏览器 等 Chromium 内核浏览器
Ver o original no GitHub

Bom para

Relacionados