Release v0.4.0 - First Public Release · ViktorTrn/codriver-mcp

CoDriver MCP v0.4.0

AI-powered Desktop Automation via Model Context Protocol

What "Claude in Chrome" is for the browser, CoDriver is for the entire desktop.

npx codriver-mcp

Or in Claude Code settings (~/.claude/settings.json):

{
  "mcpServers": {
    "codriver": {
      "command": "npx",
      "args": ["codriver-mcp"]
    }
  }
}

Tool	Description
`desktop_screenshot`	Capture full desktop, window, or specific monitor (PNG/JPEG)
`desktop_click`	Click at coordinates or by element ref
`desktop_type`	Type text at cursor or into a specific element
`desktop_key`	Press key combinations (`ctrl+c`, `alt+tab`, `f5`)
`desktop_scroll`	Scroll in any direction at position
`desktop_drag`	Drag & drop between coordinates or element refs
`desktop_windows`	List and focus application windows
`desktop_read_ui`	Read accessibility tree with ref IDs
`desktop_find`	Find UI elements by name, role, or value
`desktop_launch`	Launch, quit, or check status of applications
`desktop_ocr`	Extract text from screen via OCR (tesseract.js)
`desktop_displays`	List connected monitors for multi-display capture

Phase 1 - MVP: Screenshots, mouse, keyboard, scroll, window management
Phase 2 - Accessibility: macOS Accessibility API via JXA, element refs, natural language find
Phase 3 - Remote: HTTP/SSE transport, API-key auth, JPEG compression
Phase 4 - Polish: OCR, drag & drop, app launcher, multi-monitor support

Node.js 20+, TypeScript 5.7+ strict
@modelcontextprotocol/sdk, @jitsi/robotjs, screenshot-desktop, sharp, tesseract.js
69 tests (vitest)