7 min read

AI that controls your Android phone: Gemini, mobile-mcp, DroidRun and what each needs

In short: there are three ways to let an AI operate an Android phone. The built-in one is Gemini's screen automation, a beta that completes tasks in a handful of apps on recent Pixel and Galaxy phones, on top of a framework Google added in Android 17. The developer route uses ADB and Android's accessibility services: open-source projects such as mobile-mcp and DroidRun let any model tap, type and swipe in any app, but they need USB debugging and a computer connected to the phone. The third is companion apps that do specific things rather than driving the screen. On iPhone, apps cannot operate other apps, so only Siri, Shortcuts and developer testing tools come close.

The routes at a glance

Gemini screen automation

In the Gemini app on Android, screen automation lets Gemini carry out a multi-step task inside another app: you describe it, Gemini operates the app, and you can watch, stop it or take over at any moment. It asks for your confirmation before placing orders. It is a beta: it works with a short list of apps such as Uber, Lyft, DoorDash, Grubhub, Uber Eats and Starbucks, on supported phones including the Pixel 10 series and Samsung's Galaxy S26 series, in English and in a limited number of countries, with daily limits. Google advises against using it for sensitive information or emergencies, and you turn it on under the Gemini app's settings.

Under it, Android 17, released for Pixel phones in June 2026, adds a system framework for AI agents to automate apps with your permission; automation is blocked inside work profiles by default, companies can switch it off on managed phones, and a settings page to review which assistants can automate which apps has been spotted in development. Android also has App Functions, structured actions that apps expose to assistants instead of screen tapping, but calling other apps' functions is reserved to privileged or certified assistants, not any app you download.

mobile-mcp

mobile-mcp is an open-source MCP server that gives an assistant control of mobile devices: it reads the screen through the accessibility tree, falls back to screenshots when needed, and taps, types and swipes. It works with Android emulators and real Android phones over ADB, and with iOS simulators and, with extra developer setup, real iPhones. Clients include Claude Code, Codex, Gemini CLI, GitHub Copilot and Cursor.

What it needs: Node.js, the Android platform tools with adb, and USB debugging enabled and authorized on the phone. Good for developers who test apps or automate repetitive flows; it is not a consumer app, and the phone stays tethered to the computer, or reachable over the network through ADB.

DroidRun (Mobilerun)

DroidRun, now developed under the name Mobilerun, is an open-source framework for controlling Android, and iOS through a separate portal, with an AI agent. A small Portal app on the phone uses Android's accessibility service to read the interface; the framework on your computer talks to it through ADB, asks the model what to do next, and executes taps, swipes and typing. You choose the model provider, and an optional vision mode sends screenshots to the model. There is also a hosted version with an API and credits, and community MCP servers expose it to Claude Code and other clients.

What it needs: a computer with Python, ADB, the Portal app installed with accessibility access, Android 11 or later, and an API key for the model. Good for automating multi-step tasks across apps; expect occasional failures on unusual screens and slow app starts.

What these permissions really give away

• An accessibility service can read everything on the screen, including messages and codes, and act on any app. Install only one you trust and turn it off when you do not need it.
• USB debugging gives a connected computer deep control of the phone. Authorize only your own computer and disable it afterwards.
• Screenshots and screen text are sent to the model provider you choose, under its terms.
• Do not let an agent handle banking, two-factor codes or payments unattended: confirm those steps yourself.

Companion apps: the phone as part of a bigger setup

Often you do not need the AI to tap through your phone: you need the phone to receive and approve what an AI does elsewhere. Automation apps such as Tasker or MacroDroid run routines and can call web services. The DeskManager app for Android, our product, pairs with your Windows PC: the assistant working on the PC can send you a notification, a file or a question, ask you to approve sensitive actions, set a timer, tell you when a long job ends and, only if you allow it, read and send SMS. It does not operate other apps on the phone.

For the opposite direction, using the phone to drive a computer, see control your PC from your phone with AI.

And on iPhone?

iOS does not let one app operate the others, so there is no consumer agent that taps through your iPhone apps. What works: Siri and the Shortcuts app for automations, connected apps inside the AI assistants, and, for developers, testing tools (Xcode simulators, WebDriverAgent) that mobile-mcp and DroidRun can use. On iPhone the phone is the remote control for a computer, not the thing being controlled.

How to choose

• I have a recent Pixel or Galaxy and want to order a ride or food by voice → Gemini screen automation.
• I develop or test Android apps → mobile-mcp with Claude Code, Codex or your editor.
• I want an agent that runs multi-step tasks on my Android phone → DroidRun / Mobilerun, with a computer and ADB.
• I want my phone to receive alerts, files and approvals from an AI working on my PC → a companion app such as DeskManager's.
• I have an iPhone → Shortcuts and Siri on the phone; use the phone to drive a computer instead.

Every way to let AI use a PC, a Mac or a phone is compared in computer use agents: the complete guide.

Frequently asked questions

Can AI control my Android phone?

Yes: Gemini can complete tasks in selected apps on supported phones, and open-source tools such as mobile-mcp and DroidRun can operate any app through ADB and accessibility services, with a computer connected.

What is mobile-mcp?

An open-source MCP server that lets an AI assistant read the screen and tap, type and swipe on Android and iOS devices, real or emulated. It is aimed at developers and needs ADB for Android.

What is DroidRun?

An open-source framework, now called Mobilerun, that lets an AI agent control an Android phone through a Portal app using accessibility services and ADB, with the model of your choice.

Can Claude or ChatGPT control my phone?

Not directly from their phone apps. Developers can connect them to a phone through MCP servers such as mobile-mcp, running on a computer linked to the phone with ADB.

Can AI control an iPhone?

Not in the sense of using your apps for you: iOS does not allow it. Siri and Shortcuts cover automations, and testing tools exist for developers.

Further reading

• Computer use agents: every way to let AI control your PC, Mac or phone
• MCP server: what it is, how it works and how to use it
• Autonomous AI agents: what they are and when to keep a human in the loop