io.github.saranshbamania/mobile-device-mcp

mobile-device-mcp

MCP server that gives AI coding assistants (Claude Code, Cursor, Windsurf) the ability to see and interact with mobile devices. 49 tools for screenshots, UI inspection, touch interaction, AI-powered visual analysis, Flutter widget tree inspection, video recording, and test generation.

AI assistants can read your code but can't see your phone. This fixes that.

Why This One?

| Feature | mobile-device-mcp | mobile-next/mobile-mcp | appium/appium-mcp | |---------|:-:|:-:|:-:| | Total tools | 49 | 20 | ~15 | | Setup | npx (30 sec) | npx | Requires Appium server | | AI visual analysis | 12 tools (Claude + Gemini) | None | Vision-based finding | | Flutter widget tree | 10 tools (Dart VM Service) | None | None | | Smart element finding | 4-tier (<1ms local search) | Accessibility tree only | XPath/selectors | | Companion app (23x faster UI tree) | Yes | No | No | | Video recording | Yes | No | No | | Test script generation | TS, Python, JSON | No | Java/TestNG only | | iOS simulator support | Yes | Yes | Yes | | iOS real device | Planned | Yes | Yes | | Screenshot compression | 89% (251KB->28KB) | None | 50-80% | | Multi-provider AI | Claude + Gemini | N/A | Single provider | | Price | Free + Pro (₹499/mo) | Free | Free |

The Problem

Web developers have browser DevTools, Playwright, and Puppeteer -- AI assistants can click around, take screenshots, and verify fixes. Mobile developers? They're stuck manually screenshotting, copying logs, and describing what's on screen. They're human middleware between the AI and the device.

What This Does

Developer: "The login button doesn't work"

Without this tool:                    With this tool:
  1. Manually screenshot              1. AI calls take_screenshot -> sees the screen
  2. Paste into AI chat               2. AI calls smart_tap("login button") -> taps it
  3. AI guesses what's wrong          3. AI calls verify_screen("error message shown") -> sees result
  4. Apply fix, rebuild               4. AI calls visual_diff -> confirms fix worked
  5. Repeat 4-5 times                 5. Done.

Quick Start

Install

npx mobile-device-mcp

No global install needed. Runs directly via npx.

Prerequisites

• Node.js 18+
• Android device/emulator connected via ADB
• ADB installed (Android SDK Platform Tools)

Setup (One-time, 30 seconds)

1. Get a Google AI key (free tier available): aistudio.google.com/apikey
2. Add .mcp.json to your project root:

macOS / Linux:

{
  "mcpServers": {
    "mobile-device": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mobile-device-mcp"],
      "env": {
        "GOOGLE_API_KEY": "your-google-api-key"
      }
    }
  }
}

Windows:

{
  "mcpServers": {
    "mobile-device": {
      "type": "stdio",
      "command": "cmd",
      "args": ["/c", "npx", "-y", "mobile-device-mcp"],
      "env": {
        "GOOGLE_API_KEY": "your-google-api-key"
      }
    }
  }
}

With Pro license key (after purchasing Pro):

macOS / Linux (Pro)

{
  "mcpServers": {
    "mobile-device": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mobile-device-mcp"],
      "env": {
        "GOOGLE_API_KEY": "your-google-api-key",
        "MOBILE_MCP_LICENSE_KEY": "MDMCP-XXXXX-XXXXX-XXXXX-XXXXX"
      }
    }
  }
}

Windows (Pro)

{
  "mcpServers": {
    "mobile-device": {
      "type": "stdio",
      "command": "cmd",
      "args": ["/c", "npx", "-y", "mobile-device-mcp"],
      "env": {
        "GOOGLE_API_KEY": "your-google-api-key",
        "MOBILE_MCP_LICENSE_KEY": "MDMCP-XXXXX-XXXXX-XXXXX-XXXXX"
      }
    }
  }
}

1. Open your AI coding assistant from that directory. That's it.

The server starts and stops automatically -- you never run it manually. Your AI assistant manages it as a background process via the MCP protocol.

Verify It Works

Claude Code: type /mcp -- you should see mobile-device: Connected

Cursor: check MCP panel in settings

Then just talk to your phone:

You: "Open my app, tap the login button, type test@email.com in the email field"
AI:  [takes screenshot -> sees the screen -> smart_tap("login button") -> smart_type("email field", "test@email.com")]

You: "Find all the bugs on this screen"
AI:  [analyze_screen -> inspects layout, checks for overflow, missing labels, broken states]

You: "Navigate to settings and verify dark mode works"
AI:  [smart_tap("settings") -> take_screenshot -> smart_tap("dark mode toggle") -> visual_diff -> reports result]

No test scripts. No manual screenshots. Just describe what you want in plain English.

Works with Any AI Coding Assistant

All use the same JSON config -- just put it in the right file for your editor.

Drop Into Any Project

Copy .mcp.json into any mobile project -- Flutter, React Native, Kotlin, Swift -- and your AI assistant gets device superpowers in that directory. No global install needed.

Free vs Pro

Free (14 tools) -- no license key needed

Pro (35 additional tools) -- ₹499/mo

Get Pro License -- unlock all 49 tools. After payment, you'll receive your license key via email within 1 hour. Add it to your .mcp.json:

{
  "mcpServers": {
    "mobile-device": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mobile-device-mcp"],
      "env": {
        "GOOGLE_API_KEY": "your-google-api-key",
        "MOBILE_MCP_LICENSE_KEY": "your-license-key"
      }
    }
  }
}

AI Visual Analysis (12 tools)

Use AI vision (Claude or Gemini) to understand what's on screen.

Flutter Widget Tree (10 tools)

Connect to running Flutter apps via Dart VM Service Protocol. Maps every widget to its source code location (file:line).

iOS Simulator (4 tools)

macOS only. Control iOS simulators via xcrun simctl.

Video Recording (2 tools)

Test Generation (3 tools)

App Management (4 tools)

Performance

The server is optimized to minimize latency and AI token costs:

• 4-tier element search: companion app (instant) -> local text match (<1ms) -> cached AI -> fresh AI. smart_tap is 35x faster than naive AI calls (205ms vs 7.6s).
• Companion app: AccessibilityService-based Android app provides UI tree in 105ms (23x faster than UIAutomator's 2448ms). Auto-installs on first use.
• Screenshot compression: AI tools auto-compress to JPEG q=60, 400w -- 89% smaller (251KB -> 28KB) with zero AI quality loss.
• Parallel capture: Screenshot + UI tree fetched simultaneously via Promise.all().
• TTL caching: 5-second cache avoids redundant ADB calls for rapid-fire tool usage.

Environment Variables

Architecture

src/
|-- index.ts              # CLI entry point (auto-discovery, env config)
|-- server.ts             # MCP server factory
|-- license.ts            # License validation and tier gating
|-- types.ts              # Shared interfaces
|-- drivers/android/      # ADB driver (DeviceDriver implementation)
|   |-- adb.ts            # Low-level ADB command wrapper
|   |-- companion-client.ts # TCP client for companion app
|   +-- index.ts          # AndroidDriver class (4-strategy UI element retrieval)
|-- drivers/flutter/      # Dart VM Service driver
|   |-- index.ts          # FlutterDriver (discovery, inspection, source mapping, hot reload)
|   +-- vm-service.ts     # JSON-RPC 2.0 WebSocket client (DDS redirect handling)
|-- drivers/ios/          # iOS Simulator driver (macOS only)
|   |-- index.ts          # IOSSimulatorDriver via xcrun simctl
|   +-- simctl.ts         # Low-level simctl command wrapper
|-- tools/                # MCP tool registrations (free + pro gating)
|   |-- device-tools.ts   # Device management
|   |-- screen-tools.ts   # Screenshots & UI inspection
|   |-- interaction-tools.ts # Touch, type, keys
|   |-- app-tools.ts      # App management
|   |-- log-tools.ts      # Logcat
|   |-- ai-tools.ts       # AI-powered tools
|   |-- flutter-tools.ts  # Flutter widget inspection
|   |-- ios-tools.ts      # iOS simulator tools
|   |-- video-tools.ts    # Screen recording
|   +-- recording-tools.ts # Test generation
|-- recording/            # Test script generation
|   |-- recorder.ts       # ActionRecorder (records MCP tool calls)
|   +-- generator.ts      # TestGenerator (TypeScript/Python/JSON output)
|-- ai/                   # AI visual analysis engine
|   |-- client.ts         # Multi-provider client (Anthropic + Google)
|   |-- prompts.ts        # System prompts & UI element summarizer
|   |-- analyzer.ts       # ScreenAnalyzer orchestrator (caching, parallel capture)
|   +-- element-search.ts # Local element search (text/alias matching, no AI needed)
+-- utils/
    |-- discovery.ts      # ADB auto-discovery
    +-- image.ts          # PNG parsing, JPEG compression, bilinear resize

companion-app/            # Android companion app (Kotlin)
                          # AccessibilityService + TCP JSON-RPC for fast UI tree

Roadmap

• [ ] iOS physical device support
• [ ] Multi-device orchestration
• [ ] CI/CD integration
• [ ] Cloud device farm support

Tested On

• Devices: Pixel 8 (Android 16), Samsung Galaxy series, Android emulators
• Apps: Telegram, Instagram, Spotify, WhatsApp, YouTube, Chrome, Settings, and Flutter apps
• AI Providers: Google Gemini 2.5 Flash, Anthropic Claude
• Platforms: Windows 11, macOS (iOS simulators)
• Connection: USB and wireless ADB

License

Business Source License 1.1

• Free for individuals and non-commercial use
• Commercial use requires a paid license
• Converts to Apache 2.0 on March 23, 2030

See LICENSE for full terms.