Best Mobile MCP Servers and AI Agents for Android and iOS

The best mobile MCP server for most developers is Mobile MCP from Mobile Next, because one install drives iOS simulators, Android emulators and real iPhones and Android phones from any MCP client. Pick agent-device if you want a CLI and CI replays as well as MCP, Pinky if a person must approve risky taps, Phone Harness MCP if you have no Mac or adb setup, and Open-AutoGLM or MobileUse if you want an AI agent for Android automation rather than a tool server.

All eight picks are open source. Star counts come from the GitHub REST API as of 12 November 2026. Setup commands are quoted from each project's README and were not run for this post. AgenticMart lists these projects from public GitHub data; the protocol pages linked below have not been scanned by us.

Mobile MCP servers and agents compared

Project	Type	Android	iOS	Runs on	Stars (12 Nov 2026)	Licence	AgenticMart page
Mobile MCP	MCP server	Yes	Yes	Emulator, simulator, real devices, Mobile Next cloud	10,529	Apache-2.0	/mcp-servers/mobile-mcp
agent-device	CLI, MCP server, Node API	Yes	Yes	Emulator, simulator, real devices, device clouds	5,355	MIT	/mcp-servers/agent-device
Appium MCP	MCP server	Yes	Yes	Emulator, simulator, real devices, remote Appium	517	Apache-2.0	Not listed yet
claude-in-mobile (mcp-devices)	MCP server	Yes	Yes	Emulator, simulator, real devices	384	None detected by GitHub	/protocols/mobile-agents/claude-in-mobile
Pinky	Desktop app with MCP server	Yes (Android 10 to 17)	Yes (iOS 17.4+)	Real devices, emulator, simulator	1,988	Apache-2.0	/protocols/mobile-agents/pinky
Phone Harness MCP	Hosted MCP server	Yes	Yes	Cloud Android emulator, cloud iPhone	793	MIT	/protocols/mobile-agents/mcp
Open-AutoGLM	Phone agent and model	Yes (also HarmonyOS)	Via a separate setup guide	Real devices, emulator	26,581	Apache-2.0	/protocols/mobile-agents/open-autoglm
MobileUse	Phone agent	Yes	No	Real Android devices over ADB	267	MIT	/protocols/mobile-agents/mobile-use

An MCP server gives tools (tap, swipe, screenshot) to an agent you already use, such as Claude Code, Codex or Cursor. A phone agent ships its own loop and often its own model: you type a goal and it plans and taps by itself. If you are new to the server side, start with what an MCP server is.

1. Mobile MCP: the default mobile MCP server

Good for: testing your own app on both platforms from the coding agent you already use, without learning XCUITest or Espresso.

Mobile MCP drives apps through the native accessibility tree and, per its README, falls back to screenshots and coordinates only when it needs to. That keeps image tokens down and gives the model real element names to work with. Its platform table lists iOS Simulator, real iPhones over USB, Android emulators and real Android phones with adb.

Setup, quoted from the docs, not run for this post. The README's standard config for most clients:

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

For Claude Code without the plugin, the README gives claude mcp add mobile-mcp -- npx -y @mobilenext/mobile-mcp@latest. You need Node.js 20 or newer, Xcode command line tools for iOS and Android Platform Tools for Android.

What it can touch on the device: install and uninstall apps, clear app data, launch and terminate apps, tap, swipe, type, press hardware buttons, open URLs, set GPS location, read and replace the clipboard, record the screen, and read device logs and crash reports. Optional cloud tools can reserve a physical device on Mobile Next Cloud after a browser login.

2. agent-device: CLI, MCP and CI replays in one

Good for: teams who want an agent to check its own change in the running app, then keep that run as a repeatable test.

agent-device from Callstack is a CLI first, with a built-in stdio MCP server and a typed Node.js API on the same runtime. Its README lists iOS, Android and HarmonyOS (simulators, emulators and physical devices), plus TV platforms, web, macOS and Linux. Two features stand out for testing: runs can be saved as .ad scripts and replayed in CI, or exported as Maestro YAML, and device claims stop parallel agents in different git worktrees from grabbing the same simulator.

Setup, quoted from the docs, not run for this post:

npm install -g agent-device@latest
agent-device doctor

Then add the MCP server:

{
  "mcpServers": {
    "agent-device": {
      "command": "agent-device",
      "args": ["mcp"]
    }
  }
}

The README asks for Node.js 22.12 or newer and tells you to run doctor yourself before handing the CLI to an agent.

What it can touch: accessibility snapshots, taps, fills, scrolls, gestures, alerts, screenshots, video, logs, network data, crash details and React Native render profiles. It can also route to hosted devices on BrowserStack, AWS Device Farm, TestMu AI and Limrun.

3. Appium MCP: for teams already on Appium

Good for: QA teams with Appium capabilities files and Java test suites who want an agent in the loop.

Appium MCP lives in the Appium organisation on GitHub and wraps UiAutomator2 for Android and XCUITest for iOS. What sets it apart from the other picks is test code: its README says it can generate Java/TestNG tests from plain-language descriptions, with page object templates. It can also attach to a remote Appium server instead of running drivers locally. It is not on AgenticMart yet.

Setup, quoted from the docs, not run for this post:

{
  "mcpServers": {
    "appium-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["appium-mcp@latest"],
      "env": {
        "ANDROID_HOME": "/path/to/android/sdk",
        "CAPABILITIES_CONFIG": "/path/to/your/capabilities.json"
      }
    }
  }
}

Prerequisites per the README: Node.js 22 or newer, a JDK, the Android SDK, and Xcode on macOS for iOS. For a real iPhone, a tool called appium_prepare_ios_real_device downloads and signs WebDriverAgent.

What it can touch: Appium sessions on local or remote devices, element finding, clicks, text input and screenshots. If you run its HTTP transport, the README says to keep the listener on loopback and add authentication for anything wider.

4. claude-in-mobile (mcp-devices): Android and iOS plus debugging

Good for: developers who want UI automation and a debugger in one MCP server.

claude-in-mobile was renamed mcp-devices in version 4.0, and the all-in-one claude-in-mobile package is still maintained. Platforms are separate plugins: Android (adb), iOS (xcrun, simulator and physical devices per its iOS plugin doc), web, desktop, Aurora and HarmonyOS Next. Its debug plugin adds breakpoints and variable inspection through Android JDWP and iOS LLDB, which none of the other picks claim.

Setup, quoted from the docs, not run for this post: npm i -g claude-in-mobile, then point your client at it. The modular route is npm i -g mcp-devices, npm i -g @mcp-devices/plugin-android and mcp-devices install android. Node.js 20 or newer.

What it can touch: screenshots, taps, typing, app launch, the UI tree, a shell, logs, permissions and files, plus sensor and network simulation. GitHub detected no licence file on 12 November 2026, so check the repository before using it at work.

5. Pinky: a human approves the risky taps

Good for: letting an agent use your own real phone, where a wrong tap could send a message or a payment.

Pinky is a desktop app for macOS, Windows and Linux that exposes a phone to any MCP agent while you watch it in a window. Its README says approvals run in Pinky's gateway between the agent and the phone, so the agent cannot approve its own step. Payments, trades and transfers wait for your click by default, and you can add sends, deletions, installs and settings changes. Banking, wallet and password-manager apps are hidden from agents by default, and agents cannot type into password or one-time-code fields. The project calls itself pre-alpha.

Setup, quoted from the docs, not run for this post: install the app (brew install --cask diygod/tap/pinky on a Mac, or a release download), add the phone in the app, then use Settings › Agents › Connect, or add it by hand:

{ "mcpServers": { "pinky": { "command": "pinkyctl", "args": ["mcp"] } } }

What it can touch: Android 10 to 17 and iPhones on iOS 17.4 or later, over USB or Wi-Fi, plus simulators and emulators you created. Agents tap, type, swipe, drag, open links and change test settings such as rotation and location, which Pinky restores when the task ends.

6. Phone Harness MCP: a cloud phone with no local setup

Good for: using a phone from ChatGPT, Claude or Cursor when you have no Mac, no adb and no spare device.

Phone Harness MCP is a remote Streamable HTTP server at https://mcp.phone-harness.com/mcp with OAuth sign-in. It gives an agent a real cloud iPhone or a cloud Android emulator. The code is MIT, and the README links a self-hosting guide for stdio and Docker.

Setup, quoted from the docs, not run for this post:

claude mcp add --transport http phone-harness https://mcp.phone-harness.com/mcp

Then run /mcp and sign in.

Cost: this is the one pick with a bill. As of 12 November 2026 the README says cloud Android emulators are free to start, then 5 cents per phone-minute, and a cloud iPhone is $80 a month. Check its pricing page for current rates.

What it can touch: the cloud phone only: screenshots, element lists (accessibility tree on Android, text recognition on iPhone), taps, long presses, swipes, typing, buttons and app launch. Its tool names follow Mobile MCP's, and it has no tool for opening URLs.

7. Open-AutoGLM: an open phone agent with its own model

Good for: an AI agent for Android automation that runs from a plain-language goal, with a model you can host yourself.

Open-AutoGLM from Z.ai pairs a phone agent framework with the AutoGLM-Phone-9B models on Hugging Face. It controls Android over ADB and HarmonyOS over HDC, reads the screen with a vision-language model and plans each step. Its Chinese README links a separate iOS setup guide using WebDriverAgent. The English README describes a confirmation step for sensitive operations and a Take_over action that hands the phone back to you for logins and captchas. The last push to the repository was on 6 March 2026, so expect slower updates than the other picks.

Setup, quoted from the docs, not run for this post: pip install -r requirements.txt and pip install -e ., connect the phone, install ADB Keyboard for text input, then for example:

python main.py --base-url http://localhost:8000/v1 --model "autoglm-phone-9b-multilingual" "Open Chrome browser"

What it can touch: any app on the connected phone, through launch, tap, type, swipe, back, home, long press and double tap actions.

8. MobileUse: a research-grade Android GUI agent

Good for: researchers and tinkerers who want a multi-agent Android agent with a web UI and benchmark scripts.

MobileUse controls real Android devices over ADB using multimodal models, with reflection and planning agents on top. Its paper was accepted at NeurIPS 2025 according to the README, which also ships scripts for the AndroidWorld and AndroidLab benchmarks. It is Android only.

Setup, quoted from the docs, not run for this post: pip install mobile-use, then python -m mobile_use.webui, open http://127.0.0.1:7860 and set the model's base URL and API key.

What it can touch: whatever the connected Android phone can reach over ADB, so use a test device.

Others worth a look

The Mobile Agents directory lists more mobile AI agents open source projects, including on-device Android agents and messaging assistants. From the MCP server listings, Lamda (FIRERPA) is an Android device control platform with a built-in MCP server (8,729 stars as of 12 November 2026) that also bundles traffic capture and Frida, so it suits device farms more than simple app testing.

How we picked

We started from the 20 projects in AgenticMart's Mobile Agents category and the MCP listings returned for "mobile", "android", "ios" and "adb" by AgenticMart's search API, then added Appium MCP because QA teams search for it by name. A project made the list if it met all of these on 12 November 2026:

1. It controls a phone, emulator or simulator, not only a chat app.
2. It has an open-source licence or public code, and its README gives an install path we could quote.
3. It covers a distinct need: cross-platform testing, CI replays, Appium teams, debugging, human approval, cloud devices, or a self-contained agent.

Stars helped break ties but did not decide the order. We did not install or run any of these for this post, and nothing here is a quality or security rating. The MCP listing pages for Mobile MCP, agent-device and Lamda show a dated scan report; the protocol pages are built from public GitHub data and have not been scanned.

Before you connect a phone to an agent

These tools give a model hands on a device that may hold your messages, photos and payment apps.

• Use an emulator, a simulator or a spare test phone for anything you have not watched run.
• Read the tool list first. Install, uninstall, clear data, clipboard and shell tools are the powerful ones. On AgenticMart's MCP listing pages, check the "can touch" section.
• Keep a person in the loop for sends and payments. Pinky enforces this in its gateway; with the others it depends on your MCP client's approval prompts.
• Give the agent only what the task needs. See least privilege for AI agents.

New to adding servers to a client? Follow how to install an MCP server.

FAQ

What is a mobile MCP server?

A mobile MCP server is a Model Context Protocol server whose tools control a phone, emulator or simulator, such as taking a screenshot, listing on-screen elements or tapping. Your AI client calls those tools, and its model decides which to use for the task you describe.

Is there an iOS MCP server that works on a real iPhone?

Yes. Mobile MCP, agent-device, Appium MCP and Pinky all list real iPhones in their READMEs, and claude-in-mobile lists physical devices in its iOS plugin doc. Expect Xcode and some device preparation, such as Developer Mode or WebDriverAgent signing, while Phone Harness gives you a cloud iPhone instead.

What is the best Android MCP server?

For general app testing, Mobile MCP or agent-device, since both cover Android emulators and real phones and also work on iOS. If you only need Android and want a ready-made agent rather than tools, look at Open-AutoGLM or MobileUse.

Is there an Appium MCP server?

Yes, appium/appium-mcp in the Appium GitHub organisation, Apache-2.0, 517 stars as of 12 November 2026. It uses UiAutomator2 and XCUITest drivers and can generate Java/TestNG tests.

Can an AI agent automate Android without a computer?

Some can run on the phone itself. Pinky's phone app can do simple tasks with a model you choose, and the Mobile Agents directory lists on-device Android agents. Most testing tools here still run from a computer over adb.

Mobile MCP server or phone agent: which do I need?

Choose an MCP server if you already work in Claude Code, Codex or Cursor and want that agent to test your app. Choose a phone agent if you want to type a goal and let a dedicated model carry it out.

Built a mobile agent or MCP server?

If you maintain a mobile MCP server or phone agent, developers comparing these options need to find it. On AgenticMart you can list free now, paid selling opens soon. Submit your project and it can sit alongside the picks in Mobile Agents and the wider protocols directory.