desktop-host-mcp

A local Windows MCP server for high-permission desktop control and session-based computer-use workflows.

This project started as a thin host bridge for screenshots, input injection, OCR, UI Automation, and PowerShell. It now also includes a higher-level computer-use layer with:

• session management
• structured observation
• target finding with confidence/alignment signals
• action routing across UIA and OCR
• app profiles for notepad, settings, calculator, browser, chrome,

and edge

• browser workflow actions such as navigate, open_downloads_page, and

open_history_page

• Windows only
• PowerShell available
• Node.js 20+
• Python 3.10+

This server is intentionally high-permission.

• It runs with the effective permissions of the Windows user who launches it.
• If you launch the MCP host as Administrator, the server can interact with

elevated windows.

• desktop_run_powershell is equivalent to giving the agent a local shell with

that same permission level.

Use it only in trusted local setups.

From the repository root:

.\scripts\install-deps.ps1

This installs:

• Node dependencies via npm install
• Python dependencies via python -s -m pip install -r requirements.txt

For a direct local run:

.\scripts\start-desktop-host-mcp.ps1

Or:

npm start

See mcp-server.example.json.

Replace REPLACE_WITH_REPO_ROOT with your local checkout path.

Run:

.\scripts\test-desktop-host-mcp.ps1

Or:

npm run smoke

The smoke run writes a step-by-step summary under output/smoke-test-*.json.

• window discovery and focusing
• screenshots
• clicking, scrolling, typing, key combos
• clipboard read/write
• OCR over screen or image regions
• UI Automation probing
• local PowerShell execution
• session-b