android-mcp-server

Android MCP Server

MCP server for controlling Android emulators and devices via ADB. Gives AI assistants the ability to see, interact with, and debug Android apps — taking screenshots, tapping elements, reading logs, and documenting bugs.

npm Package | GitHub | Issues

^ Sped up for better viewing, the full demo.

Features

• 21 tools for complete Android device control
• Screenshot capture with intelligent compression (Sharp-based, max 1280px)
• UI tree inspection — read element hierarchy with bounds, text, resource IDs, and state
• Touch automation — tap, swipe, scroll, type text, press hardware keys
• Element targeting — find and tap elements by resource-id, text, or content-desc
• App lifecycle — install APKs, launch apps, inspect current activity
• Logcat integration — filter by package, log level, or timestamp
• Device management — list devices, start emulators, get device info
• Compound actions — tap_and_wait combines tap + settle + UI tree in one round trip
• Persistent ADB shell — reuses a single shell session for faster command execution
• Device info caching — queries device properties once per session
• Multi-device support — target specific devices by ID
• Zero app modifications — works with any Android app via ADB, no SDK integration needed

Prerequisites

• Node.js 18+
• Android SDK with platform-tools (ADB) and emulator
• A running Android emulator or connected device

Finding your ANDROID_HOME

The server auto-discovers the SDK at ~/Library/Android/sdk (macOS) or via ANDROID_HOME. If your SDK is elsewhere, set ANDROID_HOME in the MCP config (see below).

To check:

# macOS
ls ~/Library/Android/sdk/platform-tools/adb

# Or find it via Android Studio: Settings > Languages & Frameworks > Android SDK

Setup

Claude Code

claude mcp add --scope user android -- npx -y android-mcp-server

This registers the server globally so it's available in all projects. Use --scope project instead to limit it to the current project.

If your SDK is not in the default location:

claude mcp add --scope user --env ANDROID_HOME=/path/to/sdk android -- npx -y android-mcp-server

Claude Desktop

Add to your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "android": {
      "command": "npx",
      "args": ["-y", "android-mcp-server"],
      "env": {
        "ANDROID_HOME": "/path/to/android/sdk"
      }
    }
  }
}

VS Code

Add to your VS Code settings (.vscode/settings.json):

{
  "mcp": {
    "servers": {
      "android": {
        "command": "npx",
        "args": ["-y", "android-mcp-server"],
        "env": {
          "ANDROID_HOME": "/path/to/android/sdk"
        }
      }
    }
  }
}

Cursor

Add to your Cursor MCP config (~/.cursor/mcp.json):

{
  "mcpServers": {
    "android": {
      "command": "npx",
      "args": ["-y", "android-mcp-server"],
      "env": {
        "ANDROID_HOME": "/path/to/android/sdk"
      }
    }
  }
}

Windsurf

Add to your Windsurf MCP config (~/.codeium/windsurf/mcp_config.json):

{
  "mcpServers": {
    "android": {
      "command": "npx",
      "args": ["-y", "android-mcp-server"],
      "env": {
        "ANDROID_HOME": "/path/to/android/sdk"
      }
    }
  }
}

Project config (.mcp.json)

Add to your project's .mcp.json (checked into version control so your team gets it too):

{
  "mcpServers": {
    "android": {
      "command": "npx",
      "args": ["-y", "android-mcp-server"],
      "env": {
        "ANDROID_HOME": "/path/to/android/sdk"
      }
    }
  }
}

Build from source

git clone https://github.com/martingeidobler/android-mcp-server.git
cd android-mcp-server
npm install
npm run build
claude mcp add --scope user android -- node /path/to/android-mcp-server/dist/index.js

Available Tools

Device Management

Tool	Description
list_devices	List connected Android devices and emulators
list_avds	List available Android Virtual Devices
start_emulator	Start an AVD by name (waits up to 60s)

Screenshot & UI Analysis

Tool	Description
screenshot	Take screenshot for visual analysis. Optional save_path to save to disk
get_ui_tree	Get UI element hierarchy with bounds, text, resource IDs, and state

Interaction

Tool	Description
tap	Tap at screen coordinates
tap_element	Tap element by resource-id, text, or content-desc
tap_and_wait	Tap element, wait for UI to settle, return new UI tree — single round trip
type_text	Type text into focused input
press_key	Press key (back, home, enter, tab, delete, menu, etc.)
swipe	Swipe gesture between coordinates
scroll_to_element	Scroll until element is visible
wait_for_element	Wait for element to appear (with timeout)

Diagnostics

Tool	Description
get_logs	Get logcat output, filterable by package, log level, and time
clear_logs	Clear logcat buffer (call before reproducing a bug for clean output)
get_device_info	Get model, Android version, API level, screen size, DPI

App Management

Tool	Description
launch_app	Launch app by package name
install_apk	Install APK file
get_current_activity	Get foreground app and activity
pull_file	Pull a file from the device to local filesystem
adb_shell	Run arbitrary ADB shell command

Example Workflows

Bug documentation

"Clear the logs, open the settings screen, tap the save button, then show me the logs and a screenshot"

Claude will: clear_logs → launch_app → tap_element → get_logs(package_name="com.example.app", level="E") → screenshot(save_path="./bugs/settings-crash.png")

UI testing

"Navigate through the login flow and verify each screen matches the designs"

Claude will use screenshot + get_ui_tree to see and understand each screen, tap_element/type_text to interact, and its vision capabilities to compare against mockups or descriptions.

Smoke testing

"Install the APK, launch the app, and tap through the main screens to check nothing crashes"

Claude will: install_apk → launch_app → navigate with tap_element → get_logs(level="E") to check for errors after each screen.

How It Works

The server communicates over stdio using the Model Context Protocol. All device interaction goes through ADB — no modifications to your app are required. Screenshots are captured in memory, compressed, and returned as base64 images that the AI can see and analyze visually.

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

License

MIT - see LICENSE.