What can you do with this server?

desktop-touch-mcp is a Windows 10/11 MCP server that gives AI agents full control over a desktop environment through these capabilities: Screen Observation * screenshot — Capture the desktop, windows, or regions at multiple detail levels (meta, text, image, OCR, Set-of-Marks) and modes (diff, background, dot-by-dot) * workspace_snapshot — Full overview of all open windows with thumbnails and UI summaries * desktop_state — Lightweight read of focused window/element, cursor position, and modal state Mouse & Keyboard Input * mouse_click — Click at coordinates with auto-guard (window identity verification), double/triple-click support * mouse_drag — Drag for sliders, drag-and-drop, canvas drawing, window resizing * keyboard — Type text, send key combos (e.g. ctrl+c), or run multi-step chord sequences * scroll — Scroll by wheel, element name, or with OCR-based text reading * click_element — Click native UI elements by name or AutomationId via UIA (no coordinates needed) UI Discovery & Interaction * desktop_discover — Observe the desktop and return interactive entities with time-limited leases * desktop_act — Act on a discovered entity (click, type, drag, select) using its lease for verified targeting Browser Automation (Chrome/Edge CDP — no Selenium needed) * Open/connect, navigate, click DOM elements, fill forms (including React/Vue/Svelte), execute JavaScript, inspect forms, search elements, and get full page overviews Shell & Terminal Automation * shell_session — Background command execution with real OS process stdin/stdout/exit codes * terminal — Drive a visible terminal window (run, send input, read output) * wait_until — Server-side polling for conditions (window appears, focus changes, URL matches, terminal output contains, etc.) Window Management * Focus windows by title, pin/unpin (always-on-top), dock to screen corners, launch apps and wait for their windows Workflow Utilities * run_macro — Execute up to 50 tool calls sequentially in one MCP call (eliminates round-trip latency) * clipboard — Read/write the Windows clipboard with delivery verification * notification_show — Show Windows system tray notifications * excel — Author and run VBA macros against Excel via COM Safety & Performance * Emergency stop: Move mouse to top-left corner (within 10px of 0,0) to immediately terminate the server * Auto-guard: Verifies window identity and coordinate bounds before every action * Native Rust engine for high performance (UIA ~2ms, image diff ~0.26ms) with PowerShell fallback * Recovery signals guide re-discovery, modal dismissal, or fallback methods on errors

Which integrations are available for this server?

Provides browser automation via Chrome DevTools Protocol (CDP), enabling semantic DOM interaction, navigation, form filling, and evaluation of scripts for Brave browser. Provides visual fallback (SoM) for Electron applications that lack UI Automation support, using OCR and image analysis to detect clickable elements and enable interaction. Provides browser automation via Chrome DevTools Protocol (CDP), enabling semantic DOM interaction, navigation, form filling, and evaluation of scripts for Google Chrome. Provides visual fallback for Progressive Web Apps (PWAs) that lack UI Automation support, using OCR and image analysis to detect clickable elements and enable interaction.

How do I use desktop-touch-mcp?

1. Click on "Install Server". 2. Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state. 3. In the chat, type @ followed by the MCP server name and your instructions, e.g., "@desktop-touch-mcp Take a screenshot of my desktop and describe the open windows" That's it! The server will respond to your query, and you can continue using it as needed. Here is a step-by-step guide with screenshots.

desktop-touch-mcp

by wangkang1133

Overview Schema Related Servers Score Discussions

TypeScript

Hybrid

desktop-touch-mcp is a Windows 10/11 MCP server that gives AI agents full control over a desktop environment through these capabilities:

Screen Observation

screenshot — Capture the desktop, windows, or regions at multiple detail levels (meta, text, image, OCR, Set-of-Marks) and modes (diff, background, dot-by-dot)
workspace_snapshot — Full overview of all open windows with thumbnails and UI summaries
desktop_state — Lightweight read of focused window/element, cursor position, and modal state

Mouse & Keyboard Input

mouse_click — Click at coordinates with auto-guard (window identity verification), double/triple-click support
mouse_drag — Drag for sliders, drag-and-drop, canvas drawing, window resizing
keyboard — Type text, send key combos (e.g. ctrl+c), or run multi-step chord sequences
scroll — Scroll by wheel, element name, or with OCR-based text reading
click_element — Click native UI elements by name or AutomationId via UIA (no coordinates needed)

UI Discovery & Interaction

desktop_discover — Observe the desktop and return interactive entities with time-limited leases
desktop_act — Act on a discovered entity (click, type, drag, select) using its lease for verified targeting

Browser Automation (Chrome/Edge CDP — no Selenium needed)

Open/connect, navigate, click DOM elements, fill forms (including React/Vue/Svelte), execute JavaScript, inspect forms, search elements, and get full page overviews

Shell & Terminal Automation

shell_session — Background command execution with real OS process stdin/stdout/exit codes
terminal — Drive a visible terminal window (run, send input, read output)
wait_until — Server-side polling for conditions (window appears, focus changes, URL matches, terminal output contains, etc.)

Window Management

Focus windows by title, pin/unpin (always-on-top), dock to screen corners, launch apps and wait for their windows

Workflow Utilities

run_macro — Execute up to 50 tool calls sequentially in one MCP call (eliminates round-trip latency)
clipboard — Read/write the Windows clipboard with delivery verification
notification_show — Show Windows system tray notifications
excel — Author and run VBA macros against Excel via COM

Safety & Performance

Emergency stop: Move mouse to top-left corner (within 10px of 0,0) to immediately terminate the server
Auto-guard: Verifies window identity and coordinate bounds before every action
Native Rust engine for high performance (UIA ~2ms, image diff ~0.26ms) with PowerShell fallback
Recovery signals guide re-discovery, modal dismissal, or fallback methods on errors

desktop-touch-mcp

MCP server

Windows 10/11 专用 Computer-use MCP 服务器 — 截图、UI Automation、Chrome CDP、键鼠输入、终端 + 可靠后台 shell_session，原生 Rust 引擎 + PowerShell 回退。

快速开始

方式 A：下载 zip（推荐）

从 Releases 下载 zip
解压到任意文件夹
双击 start.bat — 完事

start.bat 首次运行自动安装依赖，之后直接启动。无需 npm/npx/构建工具。

方式 B：npx

npx -y @harusame64/desktop-touch-mcp

HTTP 模式

双击 start.bat 时加参数即可开启 HTTP 模式：

start.bat --http --port 23847 --key YOUR_KEY

注册到 Claude CLI

在 ~/.claude.json 的 mcpServers 中添加：

{
  "desktop-touch": {
    "type": "stdio",
    "command": "node",
    "args": ["C:/Tools/desktop-touch-mcp/dist/index.js"]
  }
}

默认绑定 0.0.0.0。绑定到非 localhost 地址时 API 密钥为必填，客户端须带 Authorization: Bearer <KEY> 请求头。健康检查：http://<地址>:<端口>/health。

Related MCP server: pov

核心理念

发现-操作 — desktop_discover 返回带租约的可交互实体（非原始坐标），desktop_act 操作你意图的目标而非它曾经的位置
感知防护 — 每次操作前自动验证目标窗口身份和边界，防止误窗口输入
Rust 原生加速 — UIA 焦点查询 2ms（160× 加速），图像差分 SSE2 SIMD 13~15× 加速；引擎不可用时透明回退 PowerShell

环境要求

要求	版本
操作系统	Windows 10/11（64 位）
Node.js	v20+（推荐 v22+）
PowerShell	5.1+（仅作 Rust 引擎回退）
VC++ 运行时	下载（nut-js 需要）

工具一览（32+ 个）

分类	工具	说明
🌐 主体路径	`desktop_discover`	观察桌面，返回带租约的可交互实体
	`desktop_act`	通过租约验证对实体操作（点击/输入/拖拽/选择）
👁️ 观察	`desktop_state`	轻量焦点/窗口/光标/注意信号检查
	`screenshot`	多模式捕获（text/diff/dotByDot/background）
	`screenshot_query` / `screenshot_gc`	截图缓存查询与清理
	`workspace_snapshot`	所有窗口缩略图 + UI 摘要
	`server_status`	原生引擎健康诊断
⌨️ 输入	`keyboard`	键盘输入，支持 IME 旁路
	`mouse_click` / `mouse_drag`	坐标交互 + 归位 + 强制焦点
	`scroll`	滚轮 / 定位元素 / 智能列表 / 拼接
	`click_element`	UIA 按名称/ID 点击（回退）
🌐 浏览器 CDP	`browser_open` / `browser_navigate`	幂等调试启动 + 导航
	`browser_click` / `browser_fill`	跨重绘稳定的 DOM 交互
	`browser_eval`	JS 执行 / DOM 提取 / SPA 状态
🛠️ 工作流	`shell_session`	可靠后台命令通道（进程 stdin/stdout + exit code）
	`terminal`	驱动可见控制台窗口（run/send/read）
	`wait_until`	窗口/焦点/文本/URL 状态轮询
	`window_dock` / `focus_window`	窗口吸附/置顶/聚焦
	`workspace_launch`	启动应用 + 自动检测新窗口
	`run_macro`	最多 50 个操作批量执行
	`clipboard` / `notification_show`	剪贴板 + 通知
📊 Office	`excel`	VBA 宏写入与运行

标准工作流

desktop_state          → 定向：焦点窗口/元素、模态、注意信号
desktop_discover       → 查找可操作实体（返回 lease + windows[]）
desktop_act(lease, …)  → 操作实体（返回 attention + post.perception）
desktop_state          → 确认世界按预期变化

命令执行：选对通道（重要）

场景	用哪个	原因
装软件、跑脚本、文件操作、包管理	`shell_session`	真实 OS 进程 + stdin/stdout + exit code；不依赖焦点/IME/剪贴板/SoM 日志
用户必须看见某个控制台窗口	`terminal`	驱动可见窗口（UIA/WM_CHAR），适合交互演示

# 一次性命令（推荐）
shell_session({ action: "exec", input: "whoami" })
# → ok:true 仅当 exitCode===0；非零退出返回 ok:false + stdout/stderr

# 交互式多步
shell_session({ action: "start", id: "ops", shell: "powershell" })
shell_session({ action: "write", id: "ops", input: "Get-ChildItem", waitMs: 500 })
shell_session({ action: "read",  id: "ops", clear: true })
shell_session({ action: "stop",  id: "ops" })

禁止假成功： write 只有 stdin 被 OS 接受才 ok:true；exec 只有 exitCode===0 才 ok:true。

点击优先级

browser_click(selector) — Chrome/Edge CDP（跨重绘稳定）
desktop_act(lease) — 原生/对话框/视觉（基于实体）
click_element(name | automationId) — UIA 回退
mouse_click(x, y, origin?, scale?) — 像素级最后手段

恢复提示

信号	处理
`lease_expired` / `*_mismatch` / `entity_not_found`	重新 `desktop_discover`
`modal_blocking`	`click_element(name=blockingElement.name)` 关闭后重试
`entity_outside_viewport`	`scroll(to_element)` 后重调 `desktop_discover`
`executor_failed`	回退到 `click_element` / `mouse_click` / `browser_click`

租约 TTL 自适应（上限 60s），softExpiresAtMs 约 60% 处 LLM 应考虑刷新。

截图参数速查

参数	效果	Token
`detail="image"`	PNG/WebP 像素（默认）	~443
`detail="text"`	UIA 元素 JSON + clickAt 坐标	~100-300
`detail="meta"`	仅标题+区域	~20/窗口
`dotByDot=true`	1:1 WebP，图像像素=屏幕坐标	~800
`dotByDotMaxDimension=N`	限制最长边，响应含 scale	—
`grayscale=true`	灰度，文本类减小约 50%	—
`region={x,y,w,h}`	窗口局部裁剪	—
`diffMode=true`	仅变化窗口（P帧）	~160
`ocrFallback="auto"`	UIA 稀疏时自动触发 Windows OCR	—

推荐流程：

workspace_snapshot()                     → 全面定位
screenshot(detail="text", windowTitle=X) → 获取 clickAt 坐标
mouse_click(x, y)                        → 直接点击
screenshot(diffMode=true)                → 仅检查变化

浏览器 CDP 自动化

无需 Selenium/Playwright，只需启用 Chrome 远程调试端口：

chrome.exe --remote-debugging-port=9222 --user-data-dir=C:\tmp\cdp

browser_open({launch:{}})                          → 启动 CDP Chrome + 列出标签
browser_click({selector:"#submit"})                → 查找+点击一步完成
browser_eval({action:"js", expression:"..."})      → 执行 JS
browser_fill({selector:"#email", value:"..."})     → 填充受控输入（React/Vue/Svelte 安全）
browser_navigate({url:"https://example.com"})      → CDP 导航

browser_locate 返回的坐标已含浏览器 UI 偏移和 DPI 缩放，可直接传给 mouse_click。

终端命令完成判定

terminal(action='run') 通过 until 参数控制"命令是否完成"：

模式	等待内容	适用场景
`quiet`（默认）	输出安静持续 `quietMs`	短命令
`pattern`	输出匹配预期的字符串/正则	有已知结束标记的长命令
`exit`	命令真正结束 + 返回退出码	需要完成码时

exit 模式注入与回显不同的完成标记，彻底解决哨兵误匹配问题：

terminal({
  action: 'run', windowTitle: 'pwsh',
  input: 'npm run build',
  until: { mode: 'exit', shell: 'powershell' },
})
// → { reason: 'exited', exitCode: 0 }

支持的 shell：bash、powershell。cmd.exe 尚不支持。不安全输入（未关闭引号等）直接拒绝。

鼠标归位校正

截图获取坐标后窗口可能已移动，归位系统自动校正：

层级	启用方式	功能
1	始终可用	(dx, dy) 偏移修正
2	传 `windowTitle`	窗口被遮挡时自动前置
3	传 `elementName` + `windowTitle`	UIA 重查缩放后的新坐标

mouse_click(x=500, y=300)                                      # 层级 1
mouse_click(x=500, y=300, windowTitle="记事本")                  # 层级 1+2
mouse_click(x=500, y=300, windowTitle="记事本", elementName="保存")  # 层级 1+2+3
mouse_click(x=500, y=300, homing=false)                         # 关闭归位

origin + scale 坐标换算

dotByDot 截图带 dotByDotMaxDimension 时响应含 origin 和 scale，直接传入即可自动换算：

mouse_click(x=640, y=300, origin={x:0, y:120}, scale=0.6667, windowTitle="Chrome")
# 服务器换算: screen = (0 + 640/0.6667, 120 + 300/0.6667) = (960, 570)

自动防护

操作工具传入 windowTitle 时自动防护：

✅ 验证目标窗口身份（检测进程重启/HWND 替换）
✅ 确认点击坐标在窗口矩形内
✅ 失败时返回 post.perception.status，LLM 可无截图恢复

状态	含义
`ok`	防护通过
`unguarded`	未提供 windowTitle
`target_not_found`	无匹配窗口
`identity_changed`	窗口已被替换
`unsafe_coordinates`	坐标在窗口矩形外
`needs_escalation`	需用 browser_click 或指定 windowTitle

unsafe_coordinates 或 identity_changed 时可传 fixId 批准一次性恢复（15 秒过期）。

设置 DESKTOP_TOUCH_AUTO_GUARD=0 可禁用自动防护。

强制焦点

Windows 前台保护可能阻止按键到达目标窗口。mouse_click、keyboard、terminal(send) 均支持 forceFocus: true，通过 AttachThreadInput 绕过：

{ "name": "mouse_click", "arguments": { "x": 500, "y": 300, "windowTitle": "Chrome", "forceFocus": true } }

全局默认：设 DESKTOP_TOUCH_FORCE_FOCUS=1。被拒绝时返回 ok:false + code: "ForegroundRestricted"，操作被抑制不误投。

自动停靠 CLI

MCP 启动时自动停靠承载 Claude 的终端：

{
  "mcpServers": {
    "desktop-touch": {
      "env": {
        "DESKTOP_TOUCH_DOCK_TITLE": "@parent",
        "DESKTOP_TOUCH_DOCK_CORNER": "bottom-right",
        "DESKTOP_TOUCH_DOCK_WIDTH": "480",
        "DESKTOP_TOUCH_DOCK_HEIGHT": "360"
      }
    }
  }
}

环境变量	默认值	说明
`DOCK_TITLE`	—	`@parent` 沿进程树查找终端，或用字面子串
`DOCK_CORNER`	`bottom-right`	四角可选
`DOCK_WIDTH/HEIGHT`	`480/360`	px 或比例（如 `"25%"`）
`DOCK_PIN`	`true`	置顶
`DOCK_MONITOR`	主显示器	显示器 ID
`DOCK_MARGIN`	`8`	屏幕边缘填充（px）

⚠️ 置顶窗口活动时按键会发到它而非目标。键盘操作前先 focus_window(title=...)。

截图缓存

screenshot 返回廉价的 screenshot://by-ref/{id} 链接而非内联像素，减少 token 消耗。

环境变量	默认值	说明
`SCREENSHOTS_DIR`	用户缓存目录	固定缓存路径
`SCREENSHOT_MAX_COUNT`	`200`	缓存上限
`SCREENSHOT_MAX_BYTES`	`256 MiB`	磁盘上限
`SCREENSHOT_MAX_AGE_MS`	—	超龄丢弃（opt-in）
`SCREENSHOT_AUTOPRUNE`	`on`	新增时自动清理，`0` 禁用

自动感知（Always-on）

每个 desktop_state 和 desktop_act 响应自动附加 attention 信号，操作工具传 windowTitle 时自动防护。lensId 参数在操作工具上保留供高级固定目标使用。

安全机制

紧急停止（唯一安全机制）

将鼠标移至屏幕左上角（0,0 附近 10px 以内）即立即终止 MCP 服务器。

每次工具调用前检查 checkFailsafe()
500ms 后台轮询作为长操作后备
触发半径：10px

所有按键组合和应用启动均不受限制。键盘黑名单和应用黑名单已移除。

鼠标移动速度

mouse_click、mouse_drag、scroll 均支持 speed 参数：

值	行为
省略	默认 1500 px/秒
`0`	瞬移，无动画
`1~N`	N px/秒动画

全局设置：DESKTOP_TOUCH_MOUSE_SPEED=3000。常用：0=瞬移，1500=柔和，3000=快速，5000=极速。

性能（Rust 原生引擎）

UIA 基准

函数	Rust 原生	PowerShell	加速比
`getFocusedElement`	2.2 ms	366 ms	163.9×
`getUiElements`（~60 元素）	106.5 ms	346 ms	3.3×
加权平均			~82×

图像差分（SSE2 SIMD）

函数	Rust	TypeScript	加速比
`computeChangeFraction`	0.26 ms	3.8 ms	~15×
`dHash`	0.09 ms	1.2 ms	~13×

架构

MCP 客户端
    │  stdio / HTTP
    ▼
TypeScript 服务层
    ├── Rust 原生引擎（.node 插件）
    │   ├── UIA: napi-rs + windows-rs，MTA COM 线程
    │   └── 图像: SSE2 SIMD 差分 + 感知哈希
    └── PowerShell 回退（引擎不可用时自动激活）

UI 操作层 V2

desktop_discover / desktop_act 默认开启。设 DESKTOP_TOUCH_DISABLE_FUKUWARAI_V2=1 可禁用回退到 V1。

已知限制

限制	变通方法
游戏/DirectX 全屏截图可能黑屏	用 `screenshot(mode:'background')` 或 BitBlt 回退
Chrome/WinUI3 UIA 元素为空	自动 OCR 回退，或用 `browser_open` + CDP
`browser_*` 需 Chrome 以 `--remote-debugging-port` 启动	先关 Chrome，用 `browser_open({launch:{}})`
diff 缓冲区 90s 不活动后清除	长等待后调 `workspace_snapshot` 重置
置顶窗口活动时键盘发到错误窗口	先 `focus_window`，验证 `isActive=true`
Chrome 中长破折号/智能引号被拦截	用 `use_clipboard=true`
React/Vue/Svelte 受控输入	用 `browser_fill`（原生 setter + InputEvent）

Token 成本参考

模式	Token	用途
`screenshot`（768px）	~443	一般视觉
`screenshot(dotByDot)`	~800	精确点击
`screenshot(diffMode)`	~160	操作后差异
`screenshot(detail="text")`	~100-300	UI 交互（无图像）
`workspace_snapshot`	~2000	全会话概览

环境变量汇总

变量	默认值	说明
`DESKTOP_TOUCH_API_KEY`	—	HTTP 模式 API 密钥，绑定非 localhost 时必填
`DESKTOP_TOUCH_HOST`	`0.0.0.0`	HTTP 监听地址
`DESKTOP_TOUCH_MOUSE_SPEED`	`1500`	鼠标移动速度（px/秒）
`DESKTOP_TOUCH_FORCE_FOCUS`	—	设 `1` 全局强制焦点
`DESKTOP_TOUCH_AUTO_GUARD`	`on`	设 `0` 禁用自动防护
`DESKTOP_TOUCH_DOCK_TITLE`	—	自动停靠标题（`@parent` 查终端）
`DESKTOP_TOUCH_DOCK_CORNER`	`bottom-right`	停靠角落
`DESKTOP_TOUCH_DOCK_WIDTH/HEIGHT`	`480/360`	停靠尺寸
`DESKTOP_TOUCH_DOCK_PIN`	`true`	停靠置顶
`DESKTOP_TOUCH_DOCK_MARGIN`	`8`	停靠边距
`DESKTOP_TOUCH_DISABLE_FUKUWARAI_V2`	—	设 `1` 禁用 V2 工具
`DESKTOP_TOUCH_MCP_HOME`	—	npx 缓存根目录
`DESKTOP_TOUCH_SCREENSHOTS_DIR`	用户缓存	截图缓存路径
`DESKTOP_TOUCH_SCREENSHOT_MAX_COUNT`	`200`	缓存数量上限
`DESKTOP_TOUCH_SCREENSHOT_MAX_BYTES`	`256 MiB`	缓存大小上限

许可证

MIT

Install Server

license - permissive license

quality

maintenance

How are these scores calculated?

Maintenance

–Maintainers

–Response time

3dRelease cycle

5Releases (12mo)

Commit activity

Resources

Need Help?

Related Servers

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Tools

View all tools

Related MCP Servers

OmniFlow
OS Automation App Automation
huppugo1
A
license
-
quality
B
maintenance
Enables AI clients to automate Windows desktop applications through window manipulation, image recognition, OCR, keyboard/mouse simulation, and memory operations via the MCP protocol.
Last updated 2026-06-13
MIT
pov
OS Automation Autonomous Agents
empezarcero
A
license
-
quality
D
maintenance
Enables LLM agents to capture screenshots, control mouse/keyboard, and manage windows on desktop platforms, primarily Windows, via an MCP server.
Last updated 2026-03-10
1
MIT
omni-computer-use
Autonomous Agents OS Automation
Jason26214
A
license
-
quality
A
maintenance
Enables Windows desktop automation via MCP, allowing AI agents to control mouse, keyboard, and screen capture with the same interface as Anthropic's computer-use tool.
Last updated 2026-07-23
1
MIT
native-devtools-mcp
Browser Automation OS Automation
sh3ll3x3c
A
license
-
quality
B
maintenance
Gives AI agents and MCP clients direct control over native desktop apps, Chrome/Electron browsers, and Android devices with screenshots, OCR, accessibility-based element lookup, input simulation, window management, CDP, and ADB in one local server.
Last updated 2026-05-31
120
MIT

View all related MCP servers

Related MCP Connectors

Glasswarp
Eyes and hands on real Windows PCs — observe, click, type via Glasswarp API.
Frenchie
OCR, transcription, file extraction, and image generation for AI agents via MCP.
chromedevtoolsmcp-mcp
A paid remote MCP for AI agent browser DevTools MCP, built to return verdicts, receipts, usage logs,

View all MCP Connectors

Latest Blog Posts

Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)
By Om-Shree-0709 on July 25, 2026.
mcp
Agent Identity
OAuth 2.1
Your AI Chatbot Just Exposed Your CEO's Salary to an Intern
By Om-Shree-0709 on July 2, 2026.
Agent Identity
MCP Security
OAuth Delegation
Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)
By Om-Shree-0709 on June 30, 2026.
Agentic Ai
Prompt Injection
WebAssembly

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/wangkang1133/desktop-touch-mcp'

If you have feedback or need assistance with the MCP directory API, please join our Discord server

desktop-touch-mcp

快速开始

方式 A：下载 zip（推荐）

方式 B：npx

HTTP 模式

注册到 Claude CLI

核心理念

环境要求

工具一览（32+ 个）

标准工作流

命令执行：选对通道（重要）

点击优先级

恢复提示

截图参数速查

浏览器 CDP 自动化

终端命令完成判定

鼠标归位校正

origin + scale 坐标换算

自动防护

强制焦点

自动停靠 CLI

截图缓存

自动感知（Always-on）

安全机制

紧急停止（唯一安全机制）

鼠标移动速度

性能（Rust 原生引擎）

UIA 基准

图像差分（SSE2 SIMD）

架构

UI 操作层 V2

已知限制

Token 成本参考

环境变量汇总

许可证

Maintenance

Resources

Looking for Admin?

Tools

Related MCP Servers

OmniFlow

pov

omni-computer-use

native-devtools-mcp

Related MCP Connectors

Latest Blog Posts

MCP directory API