vectora-foundry/native-devtools-mcp

View on GitHub ↗

MCP server for computer use & browser automation - screenshot, OCR, click, type, find_text, Chrome/Electron CDP, template matching. macOS, Windows & Android. Works with Claude, Cursor, and any MCP client.

126 ★15 forksRustUpdated 3mo ago

What you need to know

An MCP server for computer use on native desktop and mobile apps — macOS, Windows, Android, and Chrome/Electron via CDP — with screenshots, OCR, accessibility-first element lookup, input simulation, and window management, all in one local server.

Install

npx -y native-devtools-mcp
npm install -g native-devtools-mcp

Usage

  • Run npx native-devtools-mcp setup once to check macOS permissions and detect/write config for Claude Desktop, Claude Code, and Cursor.
  • Pick an interaction approach: Visual (screenshots + OCR) for any app, AX Dispatch for native macOS apps, or CDP for Chrome/Electron.
  • For Android, install ADB, enable USB debugging, then android_list_devices / android_connect and drive with android_* tools.
  • When the agent is driving (clicking/typing), don't move your mouse or type — real hardware inputs conflict with simulated ones.

Key features

  • Screenshots of screens/windows/regions with built-in OCR (Vision on macOS, Windows Media OCR on Windows).
  • Element-precise macOS AX Dispatch (take_ax_snapshot, ax_click, ax_set_value, ax_select) that doesn't move the cursor or steal focus.
  • Chrome DevTools Protocol automation for Chrome 136+/Chromium/Electron apps (18 CDP tools), no separate Node.js server.
  • Built-in Android support over ADB (USB or Wi-Fi): screenshots, uiautomator text lookup, input, and app management.
  • Template matching, window management, hover tracking, and recording — 100% local with no API key.

Best for

Agents that need to operate real desktop apps, Electron apps, and Android devices across macOS and Windows.

Caveats

  • No Linux support yet (macOS/Windows/Android only), and no iOS.
  • Browser automation is Chrome/Electron-only via CDP — no Firefox or WebKit (use Playwright MCP for those).
  • Headed only — depends on real-machine permissions; not a headless CI test grid.
  • macOS needs Accessibility and Screen Recording permissions, otherwise clicks silently fail and screenshots return black rectangles.
  • Claude Desktop on macOS requires the signed DMG bundle (Gatekeeper blocks npx); MIUI/HyperOS devices need an extra USB-debugging security toggle.
Platforms: macOS · Windows · Android · LocalClients: Claude Desktop · Claude Code · Cursor · VS Code · Windsurf

Reviewed 2026-08-18

Topics

Stars
126★
Forks
15
Language
Rust
License
MIT
Created
2026-01-14
Last push
2026-05-31