Mobile Device MCP
An MCP server that lets AI agents control iOS and Android devices (tap, scroll, type, take screenshots, read UI trees, and run code). Works with multiple devices at the same time.
README
Mobile Device MCP
An MCP server that lets AI agents control iOS and Android devices (tap, scroll, type, take screenshots, read UI trees, and run code). Works with multiple devices at the same time.
How It Works
Three-layer architecture:
- On-device servers — Lightweight HTTP servers running on each mobile device (UIAutomator on Android, XCUITest on iOS) that expose the accessibility tree and accept interaction commands.
- UI tree filter — Normalizes raw UI trees from both platforms into a unified flat element list.
- MCP server — The external interface. Handles device discovery, bootstrapping, port allocation, and proxies requests to on-device servers.
Devices are bootstrapped on first use — the server installs the driver app, allocates a port, starts the on-device server, and polls until it's healthy. After that, all tool calls are proxied over localhost HTTP with per-device bearer token auth.
Tools
| Tool | Description |
|---|---|
list_devices |
List available iOS and Android devices |
screenshot |
Capture the device screen (JPEG) |
uitree |
Get the UI element tree as a flat list, with optional search and limit |
tap |
Tap at screen coordinates |
double_tap |
Double-tap at screen coordinates |
long_press |
Long-press at screen coordinates (configurable duration) |
scroll |
Swipe from start to end coordinates |
type_text |
Type text into the focused element |
press_button |
Press a hardware/navigation button (home, back, enter, volumeUp/Down, dpadUp/Down/Left/Right/Center) |
launch_app |
Launch an app by bundle ID / package name |
terminate_app |
Force-stop an app |
list_apps |
List installed apps |
run_code |
Execute sandboxed JavaScript on-device (see run_code below) |
run_code
Agents can pass code that looks like UIAutomator or XCUITest, both being Javascript under the hood. The sandbox restricts (Android) potentially dangerous Java operations and only allows (iOS) some XCUITest-ish commands
- Android: Rhino engine with UIAutomator bindings —
uiDevice(click, swipe, find elements, press keys, read display info),By(selectors),Until(wait conditions),console.log() - iOS: JavaScriptCore with XCUITest bindings —
app(query elements, tap, type, swipe),springboard,device,openApp(bundleId),sleep(ms),console.log()
Both platforms automatically kill runaway scripts (infinite loops) and create a fresh sandbox per call.
Prerequisites
- Node.js 18+ (for running via
npx) - Android: Android SDK with
adbon PATH - iOS Simulator: Xcode with
xcrun,simctl - iOS Real Device: Xcode with
xcodebuild,devicectl, andiproxy(from libimobiledevice) - Building from source: Bun runtime, Gradle (Android), Xcode (iOS)
Installation
Claude Code
claude mcp add mobile-device-mcp -- npx -y @srmorete/mobile-device-mcp@latest
Or with custom ports:
claude mcp add mobile-device-mcp -e MDMS_PORT_ANDROID=20000 -e MDMS_PORT_IOS=21000 -- npx -y @srmorete/mobile-device-mcp@latest
Modifying .mcp.json (Cursor, Claude Desktop, etc)
{
"mcpServers": {
"mobile-device-mcp": {
"command": "npx",
"args": ["-y", "@srmorete/mobile-device-mcp@latest"],
"env": {
"MDMS_PORT_ANDROID": "18000", # optional
"MDMS_PORT_IOS": "19000" # optional
}
}
}
}
Building from Source
git clone <repo-url>
cd mobile-device-mcp
bun install
# Build drivers for both platforms and pack tarball
./scripts/build.sh
The build script compiles the on-device drivers (Android APKs via Gradle, iOS test bundle via xcodebuild), copies them to drivers/, and creates an npm tarball.
To run locally during development:
bun run start # Start the MCP server
bun test # Run the test suite
Configuration
| Environment Variable | Default | Description |
|---|---|---|
MDMS_PORT_ANDROID |
18000 | Base port for Android on-device servers |
MDMS_PORT_IOS |
19000 | Base port for iOS on-device servers |
Ports are assigned sequentially — first Android device gets 18000, second gets 18001, and so on. Same for iOS starting at 19000.
Acknowledgements
Mobile Device MCP server stands on the shoulders of giants such as mobile-mcp and Maestro. Used as inspiration but reframed the current approach to be multi-device and with seamless Native/WebView support (especially on Android).
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。