When you build a front end or a mobile app with Claude Code or Cursor, the most tedious step often isn’t the code. It’s opening the simulator and tapping around to check it yourself. Change one line of styling, switch over, tap a button, take a screenshot, come back, describe the problem, ask the model to fix it. A few rounds of that and your flow is gone.
Mobile MCP puts the phone directly in the agent’s hands. One set of MCP tools works against iOS simulators, Android emulators and real devices. The model can list on-screen elements, tap coordinates, type, swipe, install apps and pull logs, without you having to learn XCUITest or Espresso first. (Details checked on October 15, 2026.)
What Mobile MCP is
Mobile MCP is an open-source MCP server (MCP, the Model Context Protocol, is the standard way AI assistants plug into external tools). The repo is mobile-next/mobile-mcp, with about 9,290 stars at the time of writing, under Apache-2.0. The npm package is @mobilenext/mobile-mcp, currently at about 1.0.11, and the website is mobilenext.ai.
The core idea is to let MCP clients such as Claude Code, Cursor, Codex, Copilot and Gemini drive mobile devices through one shared toolset, instead of separate glue code for iOS and Android.
Its design choices are clear:
- Accessibility tree first: it prefers the native accessibility information to find and tap controls, which means fewer screenshots and fewer vision tokens. Screenshots plus coordinates are only the fallback.
- One API, many targets: simulators, emulators and USB-connected devices all use the same tool names.
- Device-level actions: tap, swipe, long press, install/launch/kill apps, screen recording, hardware buttons, deep links, orientation, clipboard, GPS, crash reports and more.
- Optional cloud devices: the same tools can connect to Mobile Next Cloud, so you can try it without a physical phone.
Claude Code users can also install the plugin and run /mobile-mirror to watch a live view of the device in a Ghostty or kitty side pane (there’s a demo in the README).
How to install it
Roughly, you need Node.js v20+ plus the platform tools for your target: Xcode and simctl for iOS, Android Platform Tools and adb for Android, and for real devices you’ll need to trust the computer and enable debugging.
Most clients use the same MCP config:
{
"mcpServers": {
"mobile-mcp": {
"command": "npx",
"args": ["-y", "@mobilenext/mobile-mcp@latest"]
}
}
}
In Claude Code you can also run:
claude mcp add mobile-mcp -- npx -y @mobilenext/mobile-mcp@latest
Once it’s installed, boot a simulator or plug in a device, then ask the model to call mobile_list_available_devices. There are many tools; the common ones include mobile_list_elements_on_screen, mobile_click_on_screen_at_coordinates, mobile_type_keys, mobile_take_screenshot, mobile_launch_app, mobile_get_device_logs and mobile_batch_commands. See the README or the wiki for the full list.
What I like
- It tackles a real pain point: when AI writes your app, mini-app or mobile web page, verification is still stuck on a human tapping a phone.
- Accessibility-first is more reliable and cheaper in tokens than screenshot-only approaches.
- Wide client support with a standard MCP config, so there’s little to learn.
- Apache-2.0, and local devices run on your own machine by default.
- The cloud-device option helps if you don’t have a Mac or a physical phone (it’s a separate cloud service).
What’s missing
- Setup isn’t light: Xcode, the Android SDK and device trust all have to work, and any one failing shows up as “the agent can’t find a device”.
- For complex gestures, game-style graphics or pure Canvas UIs, the accessibility tree may not be enough, so it falls back to screenshots and coordinates, and success depends on the model.
- Automating real phones touches permissions and privacy, so teams should agree on whose phone and which test accounts get used.
- The website also promotes Cloud, a CLI and Mobilewright, so newcomers can easily confuse the open-source MCP server with paid cloud minutes. Be clear about which one you’re installing.
Who it’s for
- People who use Claude Code or Cursor daily for mobile or cross-platform work and are tired of manual testing.
- Anyone who wants the agent to run through login, form filling and screenshot regression on its own.
- Developers willing to set up the simulator environment once in exchange for “describe the goal, and the machine does the tapping”.
If you rarely touch apps and mostly write back-end code, skip it. If you’re already losing half your day to the simulator, Mobile MCP is worth one install to try the full loop.
Related reading
- Braus: An Open-Source macOS Desktop Agent That Lets Claude Code and Codex Control Your Apps via MCP
- OpenCode 2.0: Local Service
Links
- Repo: https://github.com/mobile-next/mobile-mcp
- Website: https://mobilenext.ai
- npm:
@mobilenext/mobile-mcp
