Mobile Next - MCP server for Mobile Development and Automation | iOS, Android, Simulator, Emulator, and Real Devices

English | 日本語 | 简体中文

This is an MCP Server that enables scalable mobile automation, development through a platform-agnostic interface, eliminating the need for distinct iOS or Android knowledge. You can run it on emulators, simulators, and real devices (iOS and Android).

This server allows Agents and LLMs to interact with native iOS/Android applications and devices through structured accessibility snapshots or coordinate-based taps based on screenshots.

Works with Claude Code, Codex, Gemini, GitHub Copilot, Antigravity — or any MCP-compatible client.

Run it against devices on your own machine, or against real iOS and Android devices in the cloud with Mobile Next Cloud — same tools, no local setup.

https://github.com/user-attachments/assets/bb084777-beb3-4930-ae6f-8d3fe694ddde

Main use cases

How we help to scale mobile automation:

• 📲 Native app automation (iOS and Android) for testing or data-entry scenarios.
• 📝 Scripted flows and form interactions without manually controlling simulators/emulators or real devices (iPhone, Samsung, Google Pixel etc)
• 🧭 Automating multi-step user journeys driven by an LLM
• 👆 General-purpose mobile application interaction for agent-based frameworks
• 🤖 Enables agent-to-agent communication for mobile automation usecases, data extraction

Main Features

• 🚀 Accessibility-first — fast and cheap: drives apps from the native accessibility tree (no vision model, no image tokens), falling back to screenshots + coordinates only when needed.
• 📱 One API, every target: the same tools work across iOS and Android — simulators, emulators, and real devices.
• 🧠 No platform expertise required: no XCUITest, no Espresso, no per-platform glue — describe the goal and the agent does it.
• 🧰 Full device control: taps, swipes, and gestures; app install/launch/terminate; screen recording; hardware buttons; deep links; orientation.
• 📊 Structured, deterministic output: reads real UI elements and extracts structured data, cutting the ambiguity of screenshot-only approaches.

🎯 Platform Support

🔧 Available MCP Tools

Device Management

• mobile_list_available_devices - List all available devices (simulators, emulators, and real devices)
• mobile_get_screen_size - Get the screen size of the mobile device in pixels
• mobile_get_orientation - Get the current screen orientation of the device
• mobile_set_orientation - Change the screen orientation (portrait/landscape)
• mobile_set_location - Override the GPS location reported by the device, or clear the override
• mobile_clipboard - Read or replace the device clipboard

Remote Devices (Mobile Next Cloud)

• mobile_login_to_cloud_provider - Authenticate this machine with the cloud device provider (browser-based device-code login)
• mobile_list_remote_devices - List device models available to reserve from the cloud fleet
• mobile_allocate_remote_device - Reserve a physical cloud device for exclusive use
• mobile_release_remote_device - Release a reserved cloud device back to the fleet

App Management

• mobile_list_apps - List apps known by the device backend, including the current device-locale label when available
• mobile_resolve_app - Resolve an app uniquely by package name or its current device-locale label; ambiguous labels are rejected
• mobile_get_foreground_app - Get the app currently in the foreground
• mobile_get_app_capabilities - Inspect installed/enabled/launchable metadata and whether foreground verification is supported
• mobile_launch_app - Launch an app directly by package name or current device-locale label, confirming foreground by default
• mobile_wait_for_app - Wait for an app to reach the foreground and/or expose an exact visible UI target
• mobile_launch_and_ready - Launch, wait for readiness, and return a fresh frame-validated screenshot snapshot in one call
• mobile_restore_previous_app - Return to the app that was foreground before the last verified launch/link session; screen data is re-observed
• mobile_ensure_app - Verify installation, optionally install a local package, and then launch with readiness verification
• mobile_terminate_app - Stop and terminate a running app
• mobile_install_app - Install an app from file (.apk, .ipa, .app, .zip)
• mobile_uninstall_app - Uninstall an app using bundle ID or package name

Screen Interaction

• mobile_take_screenshot - Take a screenshot to understand what's on screen
• mobile_save_screenshot - Save a screenshot to a file
• mobile_list_elements_on_screen - List UI elements with their coordinates and properties
• mobile_click_on_screen_at_coordinates - Click at coordinates, a ref, or a unique exact UI target
• mobile_double_tap_on_screen - Double-tap at specific coordinates
• mobile_long_press_on_screen_at_coordinates - Long press at specific coordinates
• mobile_swipe_on_screen - Swipe in any direction (up, down, left, right)
• mobile_start_screen_recording - Start recording the device screen to a video file
• mobile_stop_screen_recording - Stop the active screen recording and save the video

Input & Navigation

• mobile_type_keys - Type text into focused elements with optional submit
• mobile_press_button - Press device buttons (HOME, BACK, VOLUME_UP/DOWN, ENTER, etc.)
• mobile_open_url - Open URLs in the device browser
• mobile_open_app_link - Bind a URL/deep link to an expected app, open it, and verify that the target app reaches the foreground

Logs & Crash Reports

• mobile_get_device_logs - Collect live device logs (logcat on Android, unified log on iOS), optionally saved to a file
• mobile_list_crashes - List crash reports available on the device
• mobile_get_crash - Get the full content of a crash report by its ID
• mobile_batch_commands - Run multiple tools in sequence in a single call (e.g. click, type, click), optionally listing screen elements at the end

🏗️ Mobile MCP Architecture

📚 Wiki page

More details in our wiki page for setup, configuration and debugging related questions.

Prerequisites

What you will need to connect MCP with your agent and mobile devices:

• Xcode command line tools
• Android Platform Tools
• node.js v20+
• MCP supported foundational models or agents, like Claude MCP, OpenAI Agent SDK, Copilot Studio

Installation and configuration

Standard config works in most of the tools:

{
  "mcpServers": {
    "mobile-mcp": {
      "command": "npx",
      "args": ["-y", "@mobilenext/mobile-mcp@latest"]
    }
  }
}

Amp

Add via the Amp VS Code extension settings screen or by updating your settings.json file:

"amp.mcpServers": {
  "mobile-mcp": {
    "command": "npx",
    "args": [
      "@mobilenext/mobile-mcp@latest"
    ]
  }
}

Amp CLI:

Run the following command in your terminal:

amp mcp add mobile-mcp -- npx @mobilenext/mobile-mcp@latest

Antigravity 2

Antigravity doesn't have a CLI command to add MCP servers, so add it manually. Edit ~/.gemini/config/mcp_config.json and add:

{
  "mcpServers": {
    "mobile-mcp": {
      "command": "npx",
      "args": ["-y", "@mobilenext/mobile-mcp@latest"]
    }
  }
}

Cline

To setup Cline, just add the json above to your MCP settings file.

More in our wiki

Claude Code

Use the Claude Code CLI to add the Mobile MCP server:

claude mcp add mobile-mcp -- npx -y @mobilenext/mobile-mcp@latest

Claude Desktop

Follow the MCP install guide, use json configuration above.

Codex

Use the Codex CLI to add the Mobile MCP server:

codex mcp add mobile-mcp npx "@mobilenext/mobile-mcp@latest"

Alternatively, create or edit the configuration file ~/.codex/config.toml and add:

[mcp_servers.mobile-mcp]
command = "npx"
args = ["@mobilenext/mobile-mcp@latest"]

For more information, see the Codex MCP documentation.

Copilot

Use the Copilot CLI to interactively add the Mobile MCP server:

/mcp add

You can edit the configuration file ~/.copilot/mcp-config.json and add:

{
  "mcpServers": {
    "mobile-mcp": {
      "type": "local",
      "command": "npx",
      "tools": [
        "*"
      ],
      "args": [
        "@mobilenext/mobile-mcp@latest"
      ]
    }
  }
}

For more information, see the Copilot CLI documentation.

Cursor

Click the button to install:

Or install manually:

Go to Cursor Settings -> MCP -> Add new MCP Server. Name to your liking, use command type with the command npx -y @mobilenext/mobile-mcp@latest. You can also verify config or add command like arguments via clicking Edit.

Gemini CLI

Use the Gemini CLI to add the Mobile MCP server:

gemini mcp add mobile-mcp npx -y @mobilenext/mobile-mcp@latest

Goose

Click the button to install:

imageInstall in Goose

Or install manually:

Go to Advanced settings -> Extensions -> Add custom extension. Name to your liking, use type STDIO, and set the command to npx -y @mobilenext/mobile-mcp@latest. Click "Add Extension".

Kiro

Follow the MCP Servers documentation. For example in .kiro/settings/mcp.json:

{
  "mcpServers": {
    "mobile-mcp": {
      "command": "npx",
      "args": [
        "@mobilenext/mobile-mcp@latest"
      ]
    }
  }
}

opencode

Follow the MCP Servers documentation. For example in ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "mobile-mcp": {
      "type": "local",
      "command": [
        "npx",
        "@mobilenext/mobile-mcp@latest"
      ],
      "enabled": true
    }
  }
}

Windsurf

Open Windsurf settings, navigate to MCP servers, and add a new server using the command type with:

npx @mobilenext/mobile-mcp@latest

Or add the standard config under mcpServers in your settings as shown above.

Read more in our wiki! 🚀

✅ Verify it works

Once the server is configured, ask your agent to list devices:

list available devices

You should get back your running simulators, emulators, and connected devices. If you do, Mobile MCP is wired up correctly. If the list is empty, make sure a simulator or emulator is running (see Prerequisites) — for more help, check the wiki.

☁️ Scale up, use a cloud device

Want to scale to hundreds of devices? Use Mobile MCP in your CI/CD pipeline?

In your Agent, prompt:

log in to mobile next cloud and then show me which remote devices are available to me

Streamable HTTP Server Mode

By default, Mobile MCP runs over stdio. To start a Streamable HTTP server instead, use the --listen flag:

npx @mobilenext/mobile-mcp@latest --listen 3000

This binds to localhost:3000. To bind to a specific interface:

npx @mobilenext/mobile-mcp@latest --listen 0.0.0.0:3000

Then configure your MCP client to connect to http://<host>:3000/mcp (or https://…/mcp behind TLS). The endpoint accepts Streamable HTTP (POST on /mcp) without an MCP session, so the transport request itself is stateless. Device lanes, screenshot snapshots, recordings, and login processes are intentionally kept in the server process; if a load balancer is used, route a device workflow to the same process (or add a shared state store before horizontal scaling). Binding to a non-loopback interface requires MOBILEMCP_AUTH.

Migration note: --listen previously served the deprecated HTTP+SSE transport on /mcp. Clients must use Streamable HTTP against http(s)://host:port/mcp. The old pure-SSE flow on /mcp is no longer available.

When binding to localhost, Host-header DNS rebinding protection is enabled automatically.

Authorization

To require Bearer token authorization on the HTTP server, set the MOBILEMCP_AUTH environment variable:

MOBILEMCP_AUTH=my-secret-token npx @mobilenext/mobile-mcp@latest --listen 3000

When set, all requests must include the header Authorization: Bearer my-secret-token. An unauthenticated server is allowed only on loopback; --listen 0.0.0.0:3000 (or another non-loopback address) fails closed unless MOBILEMCP_AUTH is set. Browser requests with an Origin header must also be listed in MOBILEMCP_ALLOWED_ORIGINS (comma-separated) unless they use the local loopback origin.

Coordinate-safe screenshot workflow

mobile_take_screenshot returns a short-lived snapshotId together with the image dimensions, input-frame metadata, and the exact server-side transform. When tapping a point read from that image, pass the same snapshotId and coordinateSpace: "screenshot" to mobile_click_on_screen_at_coordinates. The server rejects stale snapshots, changed display geometry, points outside the image, and coordinates that round outside the half-open device viewport. This avoids relying on a rounded ratio copied from prose.

When standard mobilecli does not report coordinate units or conventions on Android, the MCP reads dumpsys display, dumpsys window displays, and adb devices -l from the authorized ADB connection and cross-checks the display size, display ID, rotation, focused window, and transport. It promotes the input frame to px / edge only when the mobilecli dimensions match the current ADB display and a live transport_id is present. If ADB is unavailable, unauthorized, inconsistent, or unparseable, the frame remains unknown and coordinate taps/swipes are refused. Set MOBILEMCP_ADB_PATH to point to a specific adb executable when it is not on PATH or in the standard Android SDK location.

For structured UI output, use format: "json" with mobile_list_elements_on_screen; it returns a pure JSON array and preserves explicit false/unknown state values.

When a stable ref is not available, the click tool also accepts an exact target object such as { "label": "Continue", "type": "Button" }. The server re-reads the UI immediately before acting and refuses zero or multiple matches; it never guesses between duplicate controls.

mobile_type_keys can take the same ref or exact target to focus an input immediately before typing. This focus step is performed inside the same per-device operation lane, so a UI refresh between listing and typing is rejected instead of sending text to an unintended field.

Direct app launch and readiness

Apps can be opened without navigating the home screen. Pass a package/bundle identifier, or pass the label returned by the device backend in its current locale (設定 on a Japanese device and Settings on an English device). Ambiguous labels are rejected rather than guessed.

Prefer mobile_launch_and_ready when the next step will interact with the app. It resolves the app, launches it, verifies the foreground app, waits for optional readiness conditions, and returns a fresh frame-validated screenshot snapshot in one call.

The default launch mode is resume, which reuses the existing app task. Use the opt-in cold mode when reproducibility requires force-stopping the package first; it may discard unsaved transient state.

{
  "device": "d3befff5",
  "appName": "設定",
  "waitFor": {
    "foreground": true,
    "target": { "text": "設定" },
    "timeoutMs": 15000
  }
}

Use MOBILEMCP_APP_ALIASES for stable team vocabulary or private apps whose localized label is not exposed by a backend. The current device-locale label is resolved dynamically and is not replaced by the static aliases.

$env:MOBILEMCP_APP_ALIASES = '{"Browser":"com.android.chrome","Internal App":"com.example.internal"}'

For a shared configuration, set MOBILEMCP_APP_ALIASES_FILE to a JSON file path. Use mobile_open_app_link to bind a URL/deep link to an expected package/bundle identifier and verify that the target app becomes foreground. Custom URL schemes are rejected by default; set MOBILEMCP_ALLOW_UNSAFE_URLS=1 only when that explicit opt-in is intended.

Unknown state and cancellation behavior

If a device operation times out, is cancelled, exits abnormally, returns an invalid response, or explicitly reports an unknown result, the server never treats it as success: it quarantines that device lane. Normal follow-up operations are rejected while quarantined, preventing a later command from racing an uncertain operation. Stopping an active recording and releasing a remote device are explicit cleanup paths; they wait for in-flight work to quiesce. Only a confirmed remote release clears the old cache and quarantine state.

WebDriverAgent does not treat cancellation of the local HTTP request as proof that the device stopped. After cancellation it attempts DELETE /actions and records confirmation only when that request succeeds. Even then, the effect of the original action remains unknown, so the device stays quarantined. mobilecli 1.0.9 exposes no targeted remote-operation cancellation API, so ending the local process is reported as unconfirmed and the device is quarantined.

Clipboard values, device logs, crash bodies, stdout/stderr, response bodies, and stack traces are not copied into diagnostic logs. Tools that explicitly retrieve those values still return them to their caller; diagnostics retain only bounded error type and machine-readable failure code.

🛠️ How to Use

After adding the MCP server to your IDE/Client, you can instruct your AI assistant to use the available tools. For example, in Cursor's agent mode, you could use the prompts below to quickly validate, test and iterate on UI interactions, read information from screen, go through complex workflows. Be descriptive, straight to the point.

✨ Example Prompts

Workflows

You can specify detailed workflows in a single prompt, verify business logic, setup automations. You can go crazy:

Search for a video, comment, like and share it.

Find the video called " Beginner Recipe for Tonkotsu Ramen" by Way of
Ramen, click on like video, after liking write a comment " this was
delicious, will make it next Friday", share the video with the first
contact in your whatsapp list.

Download a successful step counter app, register, setup workout and 5-star the app

Find and Download a free "Pomodoro" app that has more than 1k stars.
Launch the app, register with my email, after registration find how to
start a pomodoro timer. When the pomodoro timer started, go back to the
app store and rate the app 5 stars, and leave a comment how useful the
app is.

Search in Substack, read, highlight, comment and save an article

Open Substack website, search for "Latest trends in AI automation 2025",
open the first article, highlight the section titled "Emerging AI trends",
and save article to reading list for later review, comment a random
paragraph summary.

Reserve a workout class, set timer

Open ClassPass, search for yoga classes tomorrow morning within 2 miles,
book the highest-rated class at 7 AM, confirm reservation,
setup a timer for the booked slot in the phone

Find a local event, setup calendar event

Open Eventbrite, search for AI startup meetup events happening this
weekend in "Austin, TX", select the most popular one, register and RSVP
yes to the event, setup a calendar event as a reminder.

Check weather forecast and send a Whatsapp/Telegram/Slack message

Open Weather app, check tomorrow's weather forecast for "Berlin", and
send the summary via Whatsapp/Telegram/Slack to contact "Lauren Trown",
thumbs up their response.

• Schedule a meeting in Zoom and share invite via email

Open Zoom app, schedule a meeting titled "AI Hackathon" for tomorrow at
10AM with a duration of 1 hour, copy the invitation link, and send it via
Gmail to contacts "team@example.com".

Running & configuration

Environment variables

Simulators, Emulators, and Real Devices

When launched, Mobile MCP can connect to:

• iOS Simulators on macOS/Linux
• Android Emulators on Linux/Windows/macOS
• iOS or Android real devices (requires proper platform tools and drivers)

Make sure you have your mobile platform SDKs (Xcode, Android SDK) installed and configured properly before running Mobile Next Mobile MCP.

Telemetry

Mobile MCP collects anonymous usage telemetry via PostHog and Scarf. To disable it, set the MOBILEMCP_DISABLE_TELEMETRY environment variable:

MOBILEMCP_DISABLE_TELEMETRY=1 npx @mobilenext/mobile-mcp@latest

For json configurations:

{
  "mcpServers": {
    "mobile-mcp": {
      "command": "npx",
      "args": ["-y", "@mobilenext/mobile-mcp@latest"],
      "env": {
        "MOBILEMCP_DISABLE_TELEMETRY": "1"
      }
    }
  }
}

Running in "headless" mode on Simulators/Emulators

When you do not have a real device connected to your machine, you can run Mobile MCP with an emulator or simulator in the background.

For example, on Android:

1. Start an emulator (avdmanager / emulator command).
2. Run Mobile MCP with the desired flags

On iOS, you'll need Xcode and to run the Simulator before using Mobile MCP with that simulator instance.

• xcrun simctl list
• xcrun simctl boot "iPhone 16"

🧩 Part of Mobile Next

Mobile MCP is one piece of a toolkit for driving real mobile devices:

• mobilewright — "Playwright for mobile." When you're ready to turn agent-driven exploration into repeatable, deterministic tests for iOS and Android, graduate to mobilewright.
• mobilecli — the universal device CLI that Mobile MCP is built on: control devices, simulators, and emulators from the command line or a JSON-RPC API.
• Mobile Next Cloud — the same stack, rented: real iOS and Android devices on demand. Just prompt your agent: log in to mobile next cloud and then show me which remote devices are available to me to get started.

🚀 Roadmap

We're continuously improving Mobile MCP. See what we're building next in ROADMAP.md — priorities are shaped heavily by community feedback, so tell us what you'd like to see.

🤝 Contributing

Contributions are welcome — code, docs, bug reports, and ideas.

• ⭐ Star the repo — the easiest way to help others discover Mobile MCP.
• Read CONTRIBUTING.md for how to build, test, and open a pull request.
• Browse open issues to find something to work on.
• Questions and ideas are also welcome in our Slack community.

Please also review our Code of Conduct.

Thanks to all contributors ❤️

We appreciate everyone who has helped improve this project.

Privacy Policy

Mobile MCP runs locally and communicates only with the devices you connect. See the Mobile Next privacy policy at https://mobilenext.ai/privacy for data collection, usage, retention, and contact information.