CC App Bridge — Automação de Apps Windows
KeithL7/cc-app-bridge
Automação de apps Windows: clique, digite, automatize via UI Automation
Ver no GitHubSkill que permite ao Claude Code controlar qualquer aplicativo desktop Windows via UI Automation, da mesma forma que controla navegadores — lançar, focar, ler interface, clicar, digitar, enviar atalhos e capturar telas. Funciona com apps nativos, Electron, Office, UWP e qualquer software que exponha a árvore UIA do Windows.
Instalação
npx convez-ai KeithL7/cc-app-bridge1 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
Casos de uso
- Automação de formulários e documentos Office
- Operações em aplicativos de chat e comunicação
- Automação de tarefas repetitivas em software de desktop
- Orquestração inteligente com fallback screenshot+coordenada para apps sem UIA
Veredito

Skill técnico e maduro que porta o padrão de automação via árvore de acessibilidade (como em navegadores) para o desktop Windows. Bem documentado, com daemon em Node + PowerShell 7, suporta 10 operações essenciais e inclui exemplos práticos. Excelente para eficiência pessoal, mas requer responsabilidade do usuário — arquitetura sólida, risco mitigado pela ausência de eval.
README
A tradução para o português deste README ainda está na fila.
CC 本地 App 点击浏览插件 · cc-app-bridge
让 Claude Code(CC)像操作浏览器一样,操作你 Windows 上的任意本地桌面应用—— 启动、聚焦、读取界面、点击、输入、按快捷键、截图。
Let Claude Code drive any local Windows desktop app the way it drives a browser — launch, focus, read the UI, click, type, hotkeys, screenshot.
这是什么
市面上的「AI 控制浏览器」桥(如 Kimi WebBridge)只能在浏览器里点网页。cc-app-bridge 把同一套架构搬到了桌面:
浏览器桥: AI → daemon → 浏览器扩展 → 网页的无障碍树(DOM)
本插件: AI → daemon → 常驻 PowerShell 引擎 → 桌面 App 的 UI Automation 树
核心洞察:浏览器的无障碍树,在 Windows 桌面上的对应物就是 UI Automation(UIA)。 所以"读页面 → 找元素 → 点击/输入"这套范式,可以一对一平移到任意桌面软件。
能干什么(示例)
只要软件把界面暴露给 Windows UIA(绝大多数原生 / Electron / Office / UWP 应用都会),就能自动化:
- 🎮 游戏收菜:识别按钮、定点点击、循环操作(无 UIA 树的纯画面游戏走「截图 + 坐标点击」兜底)
- 💬 微信回复:读取聊天列表、定位输入框、填入文字、发送
- 📱 抖音/短视频评论操作:在桌面端打开、定位评论框、输入、提交
- 📄 办公自动化:填表单、点菜单、批量处理、导出
- 🖼️ 看一眼再决策:截窗口图给 AI,让它根据画面继续操作
⚠️ 合规使用:本工具是通用 UI 自动化框架(同类如 AutoHotkey / pywinauto)。请只在你自己的账号、自己的设备上用于个人效率自动化;自动化可能违反某些平台(游戏、社交、短视频)的服务条款,请勿用于刷量、群发、垃圾评论或任何滥用行为,后果自负。
架构
AI (HTTP) ──> daemon.mjs (127.0.0.1:10087) ──NDJSON over stdio──> worker.ps1 (UIA 引擎) ──> 桌面 App
worker.ps1 ──spawns──> capture.ps1 (截图)
- daemon.mjs:Node 写的本地 HTTP 桥,
/status+/command,请求体{action,args,session}。 - worker.ps1:常驻 PowerShell 7 进程,加载 .NET UI Automation,负责读树/点击/输入/按键。
- capture.ps1:独立的一次性截图子进程(截图能力单独隔离,最小权限)。
安装
需要 Node 24+ 和 PowerShell 7(pwsh)(Win10/11 自带或一键装)。
git clone https://github.com/<your-name>/cc-app-bridge.git
cd cc-app-bridge
pwsh -File install.ps1
install.ps1 会把运行时装到 ~/.appbridge/、技能文档装到 ~/.claude/skills/appbridge/,并启动 daemon。
之后在 Claude Code 里直接说「帮我操作 XX 软件」即可触发该技能。
用法(10 个动作)
健康检查:
node "$HOME/.appbridge/bin/appbridge.mjs" status
# {"running":true,"worker_connected":true,...}
驱动桌面(任意 HTTP 客户端):
curl -s -X POST http://127.0.0.1:10087/command \
-H 'Content-Type: application/json' \
-d '{"action":"list_windows","args":{}}'
| 动作 | 作用 |
|---|---|
launch | 启动 App(之后用 find_window 绑定窗口) |
list_windows | 列出所有顶层窗口(带 @w 引用) |
find_window | 按标题绑定目标窗口到会话 |
focus | 把窗口切到前台 |
snapshot | 读 UIA 树,得到带 @e 引用的可交互元素 |
click | UIA Invoke 优先,x,y 坐标点击兜底 |
fill | 填文字(ValuePattern → 剪贴板粘贴兜底,支持中文) |
key | 发快捷键(SendKeys 语法:{ENTER}、^s…) |
screenshot | 截窗口/元素/全屏,返回文件路径 |
close_window | 关闭窗口 |
惯用法:launch → find_window 绑窗口 → snapshot 拿 @e 引用 → click/fill/key → screenshot 验证。
详见 skill/SKILL.md 与 skill/references/operations.md。
示例脚本
examples/wechat-send.mjs —— 给指定微信好友发消息(参数化:发给谁、发什么)。微信 4.0 几乎不暴露 UIA 树,所以脚本走"截图+坐标"路线,并演示了几条实战经验:从系统托盘恢复窗口、搜索联系人、点击结果、定位输入框。
默认 dry-run、安全优先(发消息是对外、不可撤回):
# 演练:搜人、开聊天、把消息打进输入框、截图,但【不发送】
node examples/wechat-send.mjs "申锦辉" "你好"
# 看过截图确认是对的人、对的内容后,再真正发送
node examples/wechat-send.mjs "申锦辉" "你好" --send
脚本每一步都输出截图路径,务必先肉眼核对"聊天对象是本人""文字已进输入框且未发送"再加 --send。
已知边界
- 仅 Windows(引擎是 Win UIA;macOS/Linux 需另写后端)。
- 没有 UIA 树的程序(部分游戏、纯 canvas/GPU 应用)只能走截图 + 坐标点击。
- 管理员权限的 App:普通权限的引擎无法向其发送输入(Windows UIPI 限制),需以管理员身份运行 daemon。
- 个别强校验"可信输入"的程序可能忽略合成的点击/按键。
- 本工具不含任意代码执行口子(刻意不做
eval),降低风险面。
License
MIT © 2026 Keith Liao
English
cc-app-bridge is a Claude Code skill that lets AI control any local Windows desktop app via Windows UI Automation (UIA) — the desktop counterpart of a browser-automation bridge. Where a browser bridge reads the DOM accessibility tree, this reads the UIA tree, so the same "read → find element → click/type" loop works on native, Electron, Office, and UWP apps.
What it can do (examples)
Any app that exposes its UI to Windows UIA can be automated:
- 🎮 Game chores / harvesting — find buttons, click, loop (pure-canvas games fall back to screenshot + coordinate clicks)
- 💬 Messaging replies (e.g. WeChat) — read the chat, locate the input box, type, send
- 📱 Short-video comment actions (e.g. Douyin/TikTok desktop) — open, locate the comment box, type, submit
- 📄 Office automation — fill forms, click menus, batch process, export
- 🖼️ Look-then-act — screenshot a window so the AI can decide the next step from pixels
⚠️ Responsible use: This is a general-purpose UI-automation framework (like AutoHotkey or pywinauto). Use it only on your own accounts and your own machine for personal productivity. Automation may violate the Terms of Service of some platforms (games, social, short-video). Do not use it for spam, mass posting, bot comments, or any abuse. You are responsible for how you use it.
Architecture
AI (HTTP) ──> daemon.mjs (127.0.0.1:10087) ──NDJSON over stdio──> worker.ps1 (UIA engine) ──> desktop apps
worker.ps1 ──spawns──> capture.ps1 (screenshot)
A Node HTTP bridge (/status, /command) in front of an always-on PowerShell 7 process that drives
.NET UI Automation. Screenshots are taken by an isolated one-shot helper (capture.ps1).
Install
Requires Node 24+ and PowerShell 7 (pwsh).
git clone https://github.com/<your-name>/cc-app-bridge.git
cd cc-app-bridge
pwsh -File install.ps1
Installs the runtime to ~/.appbridge/ and the skill to ~/.claude/skills/appbridge/, then starts the
daemon. Health check: node "$HOME/.appbridge/bin/appbridge.mjs" status.
Actions
launch · list_windows · find_window · focus · snapshot · click · fill · key · screenshot · close_window
Typical loop: launch → find_window (bind a window) → snapshot (get @e element refs) →
click/fill/key → screenshot to verify. Full reference in
skill/SKILL.md.
Example
examples/wechat-send.mjs — send a message to a named WeChat contact.
WeChat 4.0 barely exposes a UIA tree, so it uses the screenshot + coordinate path (restore from
tray, search, click result, locate input). Safe by default — it dry-runs (types but does not
send) unless you pass --send:
node examples/wechat-send.mjs "<contact>" "<message>" # dry run, screenshots for review
node examples/wechat-send.mjs "<contact>" "<message>" --send # actually send
Limitations
Windows-only (Win UIA); no-UIA apps need screenshot + coordinate clicks; can't drive elevated/admin
windows from a non-elevated worker; no arbitrary-code-execution surface (no eval, by design).
License
MIT © 2026 Keith Liao