MCP 服务器

mobile-device-mcp

MCP server that gives AI coding assistants the ability to see and interact with mobile devices. 49 tools for Android/iOS — AI-powered visual analysis (Claude + Gemini), smart tap/type by description, Flutter widget tree inspection, video recording, and test script generation. 4-tier element search with <1ms local matching. Free tier included, zero setup via npx.

README

<h1 align="center">mobile-device-mcp</h1>

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)

Get a Google AI key (free tier available): aistudio.google.com/apikey
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):

<details> <summary>macOS / Linux (Pro)</summary>

{
  "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"
      }
    }
  }
}

</details>

<details> <summary>Windows (Pro)</summary>

{
  "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"
      }
    }
  }
}

</details>

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

Tool	Config file	Docs
Claude Code	`.mcp.json` in project root	claude.ai/docs
Cursor	`.cursor/mcp.json`	cursor.com/docs
VS Code + Copilot	MCP settings	code.visualstudio.com
Windsurf	MCP settings	windsurf.com

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

Tool	What it does
`list_devices`	List all connected Android devices/emulators
`get_device_info`	Model, manufacturer, Android version, SDK level
`get_screen_size`	Screen resolution in pixels
`take_screenshot`	Capture screenshot (PNG or JPEG, configurable quality & resize)
`get_ui_elements`	Get the accessibility/UI element tree as structured JSON
`tap`	Tap at coordinates
`double_tap`	Double tap at coordinates
`long_press`	Long press at coordinates
`swipe`	Swipe between two points
`type_text`	Type text into the focused field
`press_key`	Press a key (home, back, enter, volume, etc.)
`list_apps`	List installed apps
`get_current_app`	Get the foreground app
`get_logs`	Get logcat entries with filtering

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.

Tool	What it does
`analyze_screen`	AI describes the screen: app name, screen type, interactive elements, visible text, suggestions
`find_element`	Find a UI element by description: "the login button", "email input field"
`smart_tap`	Find an element by description and tap it in one step
`smart_type`	Find an input field by description, focus it, and type text
`suggest_actions`	Plan actions to achieve a goal: "log into the app", "add item to cart"
`visual_diff`	Compare current screen with a previous screenshot -- what changed?
`extract_text`	Extract all visible text from the screen (AI-powered OCR)
`verify_screen`	Verify an assertion: "the login was successful", "error message is showing"
`wait_for_settle`	Wait until the screen stops changing
`wait_for_element`	Wait for a specific element to appear on screen
`handle_popup`	Detect and dismiss popups, dialogs, permission prompts
`fill_form`	Fill multiple form fields in one step

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).

Tool	What it does
`flutter_connect`	Discover and connect to a running Flutter app on the device
`flutter_disconnect`	Disconnect from the Flutter app and clean up resources
`flutter_get_widget_tree`	Get the full widget tree (summary or detailed)
`flutter_get_widget_details`	Get detailed properties of a specific widget by ID
`flutter_find_widget`	Search the widget tree by type, text, or description
`flutter_get_source_map`	Map every widget to its source code location (file:line:column)
`flutter_screenshot_widget`	Screenshot a specific widget in isolation
`flutter_debug_paint`	Toggle debug paint overlay (shows widget boundaries & padding)
`flutter_hot_reload`	Hot reload Flutter app (preserves state)
`flutter_hot_restart`	Hot restart Flutter app (resets state)

iOS Simulator (4 tools)

macOS only. Control iOS simulators via xcrun simctl.

Tool	What it does
`ios_list_simulators`	List available iOS simulators
`ios_boot_simulator`	Boot a simulator by name or UDID
`ios_shutdown_simulator`	Shut down a running simulator
`ios_screenshot`	Take a screenshot of a simulator

Video Recording (2 tools)

Tool	What it does
`record_screen`	Start recording the device screen
`stop_recording`	Stop recording and save the video

Test Generation (3 tools)

Tool	What it does
`start_test_recording`	Start recording your MCP tool calls
`stop_test_recording`	Stop recording and generate a test script
`get_recorded_actions`	Get recorded actions as TypeScript, Python, or JSON

App Management (4 tools)

Tool	What it does
`launch_app`	Launch an app by package name
`stop_app`	Force stop an app
`install_app`	Install an APK
`uninstall_app`	Uninstall an app

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

Variable	Description	Default
`GOOGLE_API_KEY` or `GEMINI_API_KEY`	Google API key for Gemini vision (recommended)	--
`ANTHROPIC_API_KEY`	Anthropic API key for Claude vision	--
`MOBILE_MCP_LICENSE_KEY`	License key to unlock Pro tools	--
`MCP_AI_PROVIDER`	Force AI provider: `"anthropic"` or `"google"`	Auto-detected
`MCP_AI_MODEL`	Override AI model	`gemini-2.5-flash` / `claude-sonnet-4-20250514`
`MCP_ADB_PATH`	Custom ADB binary path	Auto-discovered
`MCP_DEFAULT_DEVICE`	Default device serial	Auto-discovered
`MCP_SCREENSHOT_FORMAT`	`"png"` or `"jpeg"`	`jpeg`
`MCP_SCREENSHOT_QUALITY`	JPEG quality (1-100)	`80`
`MCP_SCREENSHOT_MAX_WIDTH`	Resize screenshots to this max width	`720`

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.