Mobile MCP: Let AI Agents Control Real iPhones and Android Devices (2026 Guide)

"Imagine if AI could just use my phone for me." With Mobile MCP, it can.

Mobile MCP (mobile-next/mobile-mcp) is an MCP server that lets AI agents — Claude Code, Cursor, Codex, Gemini CLI, and more — drive real iOS and Android devices, simulators, and emulators directly. It has 6,500+ GitHub stars, ships under Apache-2.0, and pulls 20k+ npm downloads weekly.

The key insight: it's not screenshot-vision-based. It reads the native accessibility tree — fast, cheap, and deterministic. This guide covers everything from how it works to setup and real prompts.

What you'll learn

What Mobile MCP can do (30+ MCP tools, categorized)
Setup for Claude Code / Cursor / Codex
Why accessibility-first beats screenshot-based approaches
Real prompt examples
Security notes (including its CVE history)

What Is Mobile MCP?

Item	Detail
Developer	Mobile Next
GitHub Stars	6,500+ (565 forks, Sep 2026)
License	Apache-2.0
Targets	iOS real devices & simulators, Android real devices & emulators
Protocol	MCP (Model Context Protocol)
Clients	Claude Code, Cursor, Codex, Gemini CLI, Copilot, opencode, and more

MCP is the emerging standard for connecting AI to tools. Mobile MCP adds "mobile control" as a capability: one API, every target. No XCUITest, no Espresso, no per-platform glue — the same tools work across iOS and Android, simulators and physical devices.

What Can It Do? 30+ MCP Tools

Device management: list devices, screen size/orientation, GPS override, clipboard
App management: list/launch/terminate/install/uninstall apps (.apk, .ipa)
Screen interaction: screenshots, list UI elements with coordinates, tap/double-tap/long-press, swipe, screen recording
Input & navigation: type text, press hardware buttons (HOME/BACK/volume), open URLs
Logs & crashes: collect logcat/unified log, list and fetch crash reports
Batch commands: run multiple tools in one call
Cloud devices: reserve and release real remote devices via Mobile Next Cloud

Accessibility-First: Why It's Fast and Cheap

Screenshot-based mobile AI burns thousands of image tokens per step and estimates tap coordinates from pixels. Mobile MCP reads the accessibility tree — UI elements arrive as structured data (label, coordinates, state), so the agent decides without vision. Screenshots are the fallback, not the default.

Result: near-zero image tokens, deterministic outputs, faster loops.

Setup: One Command for Claude Code

Prerequisites: Node.js v20+, Xcode CLT (for iOS), Android Platform Tools (for Android).

# Claude Code
claude mcp add mobile-mcp -- npx -y @mobilenext/mobile-mcp@latest

# Gemini CLI
gemini mcp add mobile-mcp npx -y @mobilenext/mobile-mcp@latest

Standard JSON config (works in Cursor, Windsurf, Copilot, etc.):

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

Codex (~/.codex/config.toml):

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

Verify: ask your agent "list available devices" — you should see your simulators/emulators and connected devices. Empty list? Boot one first: xcrun simctl boot "iPhone 16".

Real Prompts: One Sentence Goes a Long Way

From the official README:

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.

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.

Multi-step journeys (search → tap → type → navigate → confirm) run from a single natural-language prompt.

Advanced: SSE Server Mode for Remote Access

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

Run it on a VPS and any MCP client — local or remote — can connect to http://<host>:3000/mcp. This is how you wire mobile automation into CI/CD.

Get a VPS for your Mobile MCP server

Security Notes

GHSA-5qhv-x9j4-c3vm (fixed): mobile_open_url previously executed arbitrary Android intents (USSD codes, tel:/sms:) without scheme validation. v0.0.50+ restricts to http/https by default; other schemes need MOBILEMCP_ALLOW_UNSAFE_URLS=1.
Prompt injection: agents can be manipulated by malicious content. Keep the default URL restrictions, and prefer test devices/emulators over your personal phone.

Questions Readers Ask

Q: Is it free? Yes — Apache-2.0, fully free. Only the cloud device fleet (Mobile Next Cloud) costs money.

Q: How is this different from screenshot-based tools? Structured accessibility data instead of pixels: no image tokens, deterministic coordinates, much faster.

Q: Can I run an Android emulator on a VPS? Possible, but emulators need KVM. Look for KVM-enabled plans, or use adb against real/cloud devices.

Summary

Mobile MCP = one API for AI agents to control iOS/Android devices, 6,500+ stars
Accessibility-first: fast, cheap, deterministic
One-command setup for Claude Code; works with all major MCP clients
SSE mode enables VPS deployment and CI/CD integration

Source: mobile-next/mobile-mcp (Apache-2.0)