dsh-super-cu
The server has been validated against a running Blender desktop session, driving its UI through Windows UI Automation (element ids, invoke/set/key/click actions) — the README reports an agent building a tree model inside real Blender end-to-end, as one of its GUI automation targets.
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., "@dsh-super-cuShow the Notepad window, type 'Hello world', then save it with Ctrl+S"
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.
dsh-super-cu
Windows 电脑操作插件(DeepSeek Harness / DSH bundle) — 让智能体看得懂界面、点得准、还快。 用 UI Automation 语义文本取代截图循环:默认不截图、不猜坐标、一次调用批量做完多步。
A Windows computer-use bundle for DeepSeek Harness. Two agent tools —
cu_show(observe) andcu_do(act) — built on UI Automation with stable element ids, pattern-first actions, batched steps and a visual fallback. No dependency on screenshots for the normal path.
实测(同机对比 | |
观察延迟 p50 / p95 | 231 / 255 ms vs 1059 / 1229 ms → 快 4.57× |
文本体积(Edge 15 标签页) | 3 308 字符 vs 31 261 → 小 9.45×(~1305 vs ~9472 tokens) |
元素级稳定 id | 跨快照 100% 复用(旧方案只有坐标) |
交互元素召回率 | 100%(93/93) |
冷启动到 tools/list | 745 ms vs 6 993 ms |
工具数 | 2 个 vs 20 个 |
为什么快
窗口列表走 win32
EnumWindows(~5 ms),UIA 只进目标窗口(旧方案每次枚举整桌面 UIA ≈420 ms/次)。单窗口 DFS + 叶子剪枝 + 属性延迟读取:Edge 428 节点 walk 123 ms + 属性 65 ms。
默认纯文本,不截图:截图要多一次抓屏+压缩(~90 ms)和上千 token。
有界 DFS 早停:controls 视图最终只展示前
max_nodes个交互元素,收集满即停 —— 重窗口(436 节点、truncated)p50 265 → 135 ms(-49%),输出逐字节相同、召回 100%。批量动作:一次
cu_do完成多步(含等待与断言),把 N 个来回压成 1 个。
Related MCP server: Netwright
为什么准
稳定元素 id:UIA
RuntimeId → 进程内自增 int,同一元素跨快照 id 不变(实测复用率 100%); 失效时按(类型, 名字) → AutomationId → 祖先路径语义重定位后继续执行原动作。Pattern 优先:
invoke / setvalue / select / toggle / expand / focus走 UIA Pattern (后台生效、不抢焦点、不动鼠标),不可用才回退真实点击。歧义即拒绝:
ct:名字命中多个候选时报AmbiguousError并列出候选,绝不猜。DPI 三空间合一:启动即
SetProcessDpiAwarenessContext(PER_MONITOR_AWARE_V2), 使 UIA rect / 抓屏 / 点击落在同一物理像素空间(150% 缩放下旧做法会偏 1.5 倍)。归一化坐标:视觉兜底用
@n:0.42,0.31(0–1,相对最近截图),换算在服务端完成,模型不做算术。动作后可断言:
expect/wait在同一次调用里确认结果,而不是"做完就走"。诚实失败:
disabled控件、pattern_unavailable、剪贴板被占用、未知按键 —— 都返回ok:false+ 可操作建议,不静默成功。
工具
cu_show — 观察(默认纯文本,~50–300 ms)
win #39719780 "无标题 - 画图" 2030x1271 @264,159 [focus]
obs 226n/145 shown 239ms controls cut=budget(+61 hidden)
[6] document "开发高效精准的ComputerUse插件"
[7] button "新建会话" [i]
[15] edit "搜索会话名称" ~
[57] tabitem "对话"元素行 = [稳定 id] 类型 "名字" + 能力字母(e expand > t toggle > s selection > v value > i invoke > c focusable)+ 状态字符(* focus / + selected / - disabled / ~ offscreen)。
cut=tree 表示节点树被截断,cut=budget(+N hidden) 表示字符预算裁剪 —— 两者可同时出现。
主要参数:window(标题/进程名/#hwnd)、view(controls/text/all)、filter、max_nodes、
screenshot、region、max_image_width。
cu_do — 动作(一次调用多步)
{"steps": [
{"op": "set", "target": "#8713", "text": "关键词", "expect": "关键词"},
{"op": "key", "keys": "enter"},
{"op": "wait", "text": "搜索结果", "timeout_ms": 3000}
]}目标写法:
#id(推荐) /ct:名字/@n:ux,uy(视觉) /@img:x,y/@x,y(物理屏幕)失败即停并给一行
→建议;未执行的步骤跳过目标窗口不在前台会自动
activate(含最小化还原、托盘隐藏窗口SW_SHOW)
op 一览:
类别 | op |
鼠标 |
|
键盘 |
|
鼠标按住 |
|
UIA |
|
窗口 |
|
控制 |
|
安装
要求:Windows 10/11、Python 3.11+、DSH >= 0.2。
git clone https://github.com/suntham-t/super-cu.git
cd super-cu
./setup.ps1 # 自动探测 python、安装 Pillow + uiautomation、把绝对路径写进 cordis.patch.yml然后在 DSH 里:插件 / Plugins → 安装本地插件(install from folder)→ 选这个目录。
重新开一个会话,工具 mcp__supercu__cu_show / mcp__supercu__cu_do 就出现了。
本插件以 DSH bundle 形式分发:
package.json声明dsh.bundle.patch,cordis.patch.yml通过官方@deepseek-ai/dsh-mcp-client行拉起一个常驻 Python MCP server(stdio)。 目录结构可直接被plugin_manager(install_bundle)或插件面板安装。
插件市场 / Plugin market
本仓库按 awesome-dsh-plugin 的收录要求准备:
package.json 声明 dsh.bundle、根目录有 cordis.patch.yml、仓库带 dsh-plugin topic。
市场条目位于 .market/suntham-t__super-cu.yml,
用 ./submit-to-market.ps1 一键投稿(fork → 上传条目 → 开 PR,无需克隆 200MB 索引仓库)。
文件
package.json DSH bundle 清单(dsh.bundle.patch)
cordis.patch.yml MCP client 行(setup.ps1 写入本机路径)
setup.ps1 一次性安装:python 检测 + 依赖 + 路径替换
src/server.py MCP stdio JSON-RPC、两个工具、坐标/断言/状态
src/bridge.py UIA 快照 / 稳定 id / 解析 / 动作 / 截图 / wait
src/windows.py win32 窗口枚举、激活升级链、剪贴板、DPI
src/compact.py 紧凑格式化(去噪 / 折叠 / 预算 / cut 标记)
src/dpi.py DPI 三空间合一
src/gui_lock.py 多智能体共享桌面的互斥锁
src/mock_bridge.py 离线 mock(协议层测试用)
tests/ 单测 + 真机集成测试(GUI 部分)
SKILL.md 给智能体的操作手册
docs/CONTRACT.md 完整接口契约(实现细节)
docs/images/ 真实任务截图真实验证(不只是单测)
基准:26 个应用/交互场景矩阵(16 PASS / 1 FAIL→已修 / 9 SKIP)、7 组压力与鲁棒性 (20 步压力 RSS +0.05 MB / 句柄 +0;并发锁 ratio 0.66;
@n误差 worst 0.71 px;50 次 id 解析 0 失败)、 输出质量审计(影子 text 残留 0)。端到端:记事本输入并保存到磁盘(内容逐字比对)、计算器 7×8=56、PowerPoint 6 页 PPT 并发送到微信 (无 API 客户端)、微信发消息、画图里画笑脸、Blender 建一棵树、在真实 Minecraft 里用
/fill盖了一栋 7×7 带门/窗/屋顶的房子。回归:
tests/test_bridge.py29 项(含剪贴板、拖拽、UWP 重定位、修饰键、hold)+tests/test_server.py141 项 mock 协议测试。
|
|
|
|
已知限制
无 UIA 的界面(canvas / 游戏 / 部分自绘 UI):只能用
screenshot=true+@n坐标,慢且依赖视觉。Chromium 的
view=text取不到正文(TextPattern 为空)→ 用controls(TextControl 的 name 就是文本)。控制台窗口(conhost)UIA 只有十几个节点、读不到输出文本 → 退路:截图或用非 UI 手段。
UIPI:管理员权限窗口可能拒绝输入/读取。
指针锁定 + raw input 的游戏(如 Minecraft):
SetCursorPos无法转视角;要用游戏内命令或键位。Office 现代「保存此文件」对话框不接受全路径 → 先「更多选项…」→「浏览」进经典对话框。
一次
cu_do建议 ≤10 步(批量期间持有 GUI 锁,其它智能体会等待)。
License
MIT © 2026 suntham-t
English TL;DR
dsh-super-cu is a DeepSeek Harness bundle that gives an agent two tools — cu_show (a compact,
token-cheap accessibility snapshot of one window with stable element ids) and cu_do (a batched
action runner with waits/assertions). Compared with screenshot-based computer use it is 4.6× faster,
returns 9.5× less text and reuses element ids across snapshots (100%). Accuracy comes from
UI-Automation patterns, refusing ambiguous targets, DPI-space unification, normalized fallback
coordinates and loud failures. Works on Windows 10/11 with Python 3.11+; install with setup.ps1
then add the folder in DSH's plugin manager.
This server cannot be deployed
Maintenance
Related MCP Connectors
Eyes and hands on real Windows PCs — observe, click, type via Glasswarp API.
- TabfleetOAuthcom.tabfleet
Launch, inspect, control, and share isolated cloud browsers for your agents.
Securely control computers you explicitly pair through files, terminals, processes, screenshots, desktop UI/input, clipboard, browser automation, diagnostics, and document tools.
Run multi-step tasks in a real Chrome browser: persistent environments, live view, human takeover.
Related MCP Servers
- FlicenseBqualityCmaintenanceEnables local Windows UI automation and screen capture, targeting windows that are difficult to automate such as games and legacy apps, with tools for window management, mouse and keyboard input, and desktop capture.8-
- AlicenseNot gradedqualityAmaintenanceEnables AI agents to launch, inspect, and operate WPF, WinForms, and WinUI desktop applications through UI Automation, with compact snapshots, background actions, state verification, and export to regression tests.18MIT
- AlicenseNot gradedqualityAmaintenanceEnables Windows agents to control desktop applications and isolated Chromium sessions through a unified computer tool, supporting discovery, observation, input, window management, browser automation, and sequencing actions.570 npm1MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to inspect and drive native Windows desktop applications (WinForms, WPF, WinUI 3, and WebView2 hybrids) through Microsoft UI Automation by capturing interactive control snapshots, then clicking, filling, and asserting on elements via ephemeral refs or semantic screen-map targets. Mutating actions are dry-run by default and gated behind explicit arming, allow-lists, and a kill switch for lawful, assisted QA testing.MIT



