Deskhand-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Deskhand-mcpopen Notepad, type 'Hello, world!', and read it back"
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.
Deskhand-mcp
Windows 桌面操控 MCP server——DeskHand 引擎的独立发行版,采用 host-as-VL 架构: 本 server 只做确定性的事(UIA 控件树 / 元素记忆 / OCR / 像素 diff / 精确读写), 一切视觉判断交回主模型——标注图通过 MCP ImageContent 直接进主模型上下文, 不需要额外的视觉模型。
从 DeskHand(AstrBot 插件版)到 Deskhand-mcp 的核心变化:
DeskHand | Deskhand-mcp | |
定位漏斗 | 记忆 → OCR → VL 三级(需配 VL 模型) | 确定性两级:记忆 → OCR;找不到就把图给主模型看 |
hover-verify | 调 VL 确认落点 |
|
精确读取 | OCR(中文混合串实测准确率约 63%) |
|
文本写入 |
|
|
运行环境 | AstrBot 插件 | 任意 MCP host(Claude Code / Kimi Code / 等) |
安装
pip install deskhand-mcp[all] # 或 pip install -e .[all]
deskhand-mcp # stdio server,直接配进 MCP host依赖说明:mcp/pywin32/Pillow/uiautomation 为必需;winsdk(WinRT OCR)、
rapidocr-onnxruntime(OCR 兜底)、opencv-python-headless(CV 候选框)、
windows-capture(WGC 后台抓图)可选,缺哪个对应通道就静默降级。
Related MCP server: opencode-gui-bridge
工具一览(14 个)
看界面(图随结果返回,主模型自己看图)
look(window="")— 编号元素卡片(UIA 控件 + OCR 文字 + CV 候选框三通道)元素标注图(框和编号与卡片一一对应)
scene_grid(window="")— 图形/自绘界面的 3×3 编号网格图 + 格→屏幕坐标映射verify_point(x, y)— 画十字准星返回 320px 局部放大图,主模型看图确认落点
操作
click(element="eN" | target="文字" | x, y)— UIA 控件后台点击不抢焦点; element 路径带现场签名校验 + OCR 自愈;动作后像素 diff 验证type_text(text, element=)— UIA 输入控件走后台安全追加(回读验证), 否则点击聚焦 + 键鼠输入press_key / scroll / drag / wait_change / window_action— 按键 / 滚动 / 拖拽 / 等画面变化 / 窗口管理(min/max/restore/close/focus/move/resize…)
精确读写(ValuePattern / TextPattern,逐字精确)
read_value(element=)— 读输入框当前值(ValuePattern)read_document(element=)— 读文档全文(TextPattern → ValuePattern → OCR 降级, OCR 结果如实标exact:false)append_text(text, element=)— 安全追加:读旧值 → 身份校验 → 追加 → SetValue → 回读逐字验证set_value(text, confirm_overwrite=true)— 整篇覆盖(危险,默认拒绝)
在 MCP host 中配置
{
"mcpServers": {
"deskhand": {
"command": "deskhand-mcp"
}
}
}未安装为包时可直接用源码运行:
{
"mcpServers": {
"deskhand": {
"command": "python",
"args": ["-m", "deskhand_mcp.server"],
"cwd": "C:\\path\\to\\deskhand-mcp"
}
}
}记忆库(元素位置 + 图像签名)存于 %LOCALAPPDATA%\deskhand-mcp\deskhand_memory.db,
设环境变量 DESKHAND_DATA_DIR 可改目录。
架构:为什么砍掉 VL
实测对比(见 DeskHand 深度评测):VL 漏斗最贵的一环(每点一次都要调视觉模型) 恰恰是最不可靠的一环——黑盒坐标、不可解释、延迟高。而现代主模型基本都已多模态, 看图能力远超专用小 VL。host-as-VL 的思路:把「看」还给主模型,把「做」做扎实——
能确定性的绝不给模型猜:UIA 控件树、OCR 文字、像素 diff、ValuePattern 回读;
必须视觉判断的地方,把图直接塞进主模型上下文(标注图/网格图/准星放大图), 由主模型下结论或给修正量——agent-in-the-loop,每一步都可解释;
精确读写绕过像素层:TextPattern 的
DocumentRange.GetText(-1)读全文逐字精确, 彻底告别「看图识字」的保真度问题。
安全边界
UAC 提权窗口、以管理员运行的窗口:UIA/win32 无法穿透,工具会如实报错;
最小化窗口 UIA 树为空:先
window_action(action="restore");set_value覆盖写入默认拒绝,必须显式confirm_overwrite=true;目标身份不符(弹窗遮挡/界面移动)时报 stale 中止,绝不盲点旧坐标。
协议
AGPL-3.0-only,与 DeskHand 一致。
This server cannot be deployed
Maintenance
Related MCP Connectors
Eyes and hands on real Windows PCs — observe, click, type via Glasswarp API.
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Hosted real Google Chrome MCP with per-user persistent state. Navigate, click, type, screenshot.
Melaya is a remote MCP server. It gives an assistant hands on your own Android phone and browser: it reads the screen through the accessibility tree, then taps, types and navigates inside the apps and sites you allow-list, with no per-app API. It also builds, schedules and runs agent pipelines across 6k+ connected tools. OAuth 2.1, nothing to install.
Related MCP Servers
- AlicenseAqualityCmaintenanceAllows AI clients to see and control Windows 10/11 desktops via MCP, with screenshots, UI Automation, Chrome CDP, keyboard/mouse, and terminal using semantic element targeting.30649 npmMIT
- FlicenseNot gradedqualityBmaintenanceEnables AI assistants to control Windows GUI by listing and focusing windows, capturing element snapshots via UIA/OCR/CDP, performing clicks/inputs/scrolls, verifying changes, waiting for screen updates, taking screenshots, and obtaining visual descriptions.2-
- AlicenseNot gradedqualityBmaintenanceEnables AI clients on Windows to control the local mouse, keyboard, and screen understanding via MCP stdio, allowing automated workflows like viewing the screen, locating elements, clicking, typing, and verifying results.1Apache 2.0
- FlicenseNot gradedqualityCmaintenanceEnables MCP clients to observe and control Windows desktop applications by reading UIA element trees and capturing screenshots, then acting through UIA patterns or synthesized mouse and keyboard input. It runs standalone on an interactive Windows session, without requiring the original host application.1-