Yu-tao-Li/dsh-computer-use-win

中文

dsh-computer-use-win

Windows computer use for DeepSeek Harness — an MCP stdio server backed by a PowerShell + UI Automation engine. It lets the agent see the desktop (accessibility tree, window-cropped screenshots, OCR with word boxes) and act on it (mouse, keyboard, semantic UIA actions, window management) across 22 tools.

Bridges into DSH through the in-box @deepseek-ai/dsh-mcp-client — no DSH modifications. Zero runtime dependencies.

① File Explorer: find locates the file, semantic invoke selects it (status bar: "1 item selected")	② type_text writing a real document (clipboard path, verifyValue read-back)	③ UWP Calculator: find locates buttons, semantic Invoke clicks (25×4=100)

Features

• Text-first observation — UIA accessibility tree in three views (control / content / raw); find controls by name, automation id, class or value instead of guessing pixels. Element ids use UIA RuntimeIds (stable across UI refresh; stale ids report stale instead of mis-clicking).
• Three-level screenshot chain — PrintWindow (works on non-foreground windows) → WGC (Windows.Graphics.Capture, for DirectComposition/occluded windows) → screen-region fallback with an occludedPossible flag. Window crop + downscale + imageScale/origin coordinate mapping; PNGs auto-GC after 30 min.
• Three input paths, honestly reported — semantic (UIA Invoke/Toggle/Value patterns, no mouse), foreground (SetCursorPos + SendInput + clipboard paste / KEYEVENTF_UNICODE), background (PostMessage into the window queue, no foreground steal; delivery is unverified and reported as verified:false — Chromium/Electron/WinUI silently drop synthetic messages, and the tool tells you).
• OCR fallback for UIA-blind apps — Windows.Media.Ocr via a csc-compiled C# WinRT helper; returns text plus per-word screen-coordinate boxes, and upgrades a query match to the control under the matched word (ControlFromPoint).
• Safety built in
  • Coordinate homing — window moves after observation are compensated; the click reports homed:{dx,dy}.
  • Failsafe panic brake — park the physical mouse in the screen corner for 500 ms and all input is refused (EMERGENCY STOP); only a human moving the mouse releases it.
  • Identity guard — acting on a windowTitle whose HWND/PID changed since last observation is refused (identity_changed).
  • Foreground verification — typed input is fail-closed: if the target window is not truly foreground, the action errors instead of hitting the wrong app.
  • Win-key blacklist — Win+R/X/L/S style chords are refused by design.
• Persistent backend — one long-lived PowerShell process with a JSON-line protocol; hot latency 6–100 ms (first call ~500 ms cold). C# P/Invoke helpers compile once to hash-named cached DLLs.

Install

# from GitHub (choose your profile)
dsh plugin --profile web add github:Yu-tao-Li/dsh-computer-use-win
# or from the dshmarket plugin market: search "dsh-computer-use-win"

Restart dsh web. The tools appear as mcp__wincu__windows_computer_use_*.

The bundle resolves its own paths at install time (cordis.patch.yml uses !!js relative to the package dir), so it installs cleanly into any profile / $DSH_HOME — no hardcoded paths. Keep only one serverName: wincu row per profile.

Tools (22, prefix mcp__wincu__)

Group	Tools
Observe	health · snapshot (tree + screenshot) · accessibility_tree · list_windows · find · element_info · ocr
Act	click · double_click · move · drag · scroll · type_text (clipboard / sendinput / background) · keypress · focus · invoke · set_value
Window	activate_window · move_window · close_window (WM_CLOSE; the app may veto with a save dialog) · wait_for · wait

Every tool accepts optional window targeting (windowTitle substring / processId / nativeWindowHandle) plus activate: true; untargeted calls act on the foreground window.

Architecture

DSH agent
   │  sees mcp__wincu__windows_computer_use_* tools
   ▼
@deepseek-ai/dsh-mcp-client   ← DSH in-box bridge (official MCP SDK, StdioClientTransport)
   │  JSON-RPC 2.0 over stdio (newline-framed)
   ▼
mcp/server.mjs                ← this repo: MCP server (Node ≥ 22, zero deps)
   │  spawns one persistent backend; JSON-line stdio protocol
   ▼
scripts/windows-uia.ps1       ← this repo: desktop engine (PowerShell 5.1)
   ├─ UI Automation assemblies        → accessibility tree
   ├─ user32 P/Invoke (cached C# DLL) → mouse / keys / window ops
   ├─ Windows.Media.Ocr (C# WinRT)    → OCR
   └─ Windows.Graphics.Capture        → WGC window capture

Safety & limitations

• Actions are real. Have the agent snapshot before sensitive operations; scope important windows with windowTitle.
• WinUI / Chromium / Electron drop PostMessage synthetic input (verified:false is honest reporting, not success). Use the foreground/clipboard paths for them.
• Elevated windows (UAC/admin) are not readable or clickable — Windows blocks both UIA and synthetic input.
• Coordinates are physical pixels (explicit Per-Monitor V2 DPI).
• OCR word boxes for CJK text are per-character (engine behavior) — fine for clicking.
• Windows only.

Development

mcp/server.mjs        MCP stdio server + persistent backend manager
scripts/windows-uia.ps1  desktop engine (one-shot & -Persistent modes)
test/                 self-test / MCP protocol / Notepad E2E / features / benchmarks
docs/wiki/            upstream architecture docs (see THIRD_PARTY.md)
docs/dev-notes.md     design rationale, pitfalls, benchmarks, test log

node mcp/server.mjs --self-test          # full stack smoke test (Windows only)
node test/mcp-test.mjs                   # initialize → tools/list → tools/call
node test/notepad-e2e.mjs                # real input E2E (opens Notepad)

CI (.github/workflows/ci.yml) runs the self-test + protocol test on windows-latest for every push/PR.

License

MIT — see LICENSE. This project is a derivative of cgissing/windows-computer-use (MIT); upstream copyright is preserved and detailed in THIRD_PARTY.md.