AI Agents for Android: Apps That Actually Operate Your Phone
November 16, 2026

The most capable mainstream AI agent that operates third-party Android apps for you is Gemini, through a beta feature Google calls multi-step tasks. It works on a short list of recent phones (Pixel 10 and 11, Galaxy S26, Z Fold 8 and Z Flip 8, and Tab S12), in five countries, and only in select apps for rides, food, groceries and, on newer devices, shopping, travel and tickets. Everything else that "controls your phone" falls into one of two camps: assistants that change settings and reply to notifications, or developer frameworks like Mobilerun (formerly DroidRun) and mobile-use that drive a phone over adb or an accessibility service. Those frameworks are impressive, but they're tools for people comfortable with a terminal.
Here's what's real today:
- Gemini multi-step tasks run the app in a secure virtual window and finish tasks in the background. You confirm the final step.
- Gemini device assistance handles alarms, timers, opening apps, notifications, media and settings.
- Galaxy AI adds Now Brief, Now Nudge and an upgraded Bixby, while Gemini does the cross-app tasks on Samsung's newest phones.
- Open-source agents (Mobilerun, mobile-use) and MCP servers (Mobile MCP) let you control a phone from your computer with your own model.
- Google Play bans autonomous agents built on the Accessibility API, which is why you won't find a polished "AI that uses your apps" app in the store.
New to AI on phones in general? Start with our guide on how to use AI on your phone. This one is about agents that take actions.
Quick Comparison
| Option | Who it's for | How it controls apps | App coverage | Where it runs |
|---|---|---|---|---|
| Gemini multi-step tasks | Everyday users on supported phones | Screen automation in a virtual window | Select apps in a few categories | On the phone, beta |
| Gemini device assistance | Anyone using Gemini on Android | System actions | Settings, alarms, notifications, media | On the phone |
| Galaxy AI and Bixby | Samsung owners | Suggestions and device help | Samsung apps and settings | On the phone |
| Mobilerun (DroidRun) | Developers, QA teams | Accessibility tree plus screenshots, via a Portal app or adb | Any app it can see | Your computer or Mobilerun Cloud |
| mobile-use | Developers | UI automation from natural language | Any app with accessibility data | Your computer |
| Mobile MCP | Coding-agent users | MCP tools over accessibility snapshots or taps | Any app on a connected device | Your computer |
Gemini on Android: What It Can Actually Do
Gemini has two different powers on Android, and they're often confused.
Multi-step tasks in other apps
Google announced the feature in February 2026 as an early preview of multi-step tasks. You long-press the power button and ask something like "book me a ride home" or "reorder my last meal on DoorDash". Gemini works in the background, and you can follow along through notifications.
Google's Pixel help page for the feature gives the current rules:
- Devices: Pixel 10, 10 Pro, 10 Pro XL, Pixel 11, 11 Pro, 11 Pro XL and 11 Pro Fold. Samsung Galaxy S26 series, Z Flip 8, Z Fold 8 and Tab S12 series.
- Countries: India, Japan, South Korea, the UK and the US. Pixel 10 isn't supported in South Korea.
- Account: 18 or older, signed in with a personal Google Account. Work, school and supervised accounts are excluded.
- Languages: select apps in English, Japanese and Korean.
- Tasks: rides, food and groceries on every supported device. Starting with the Z Flip 8 and Z Fold 8, also buying and selling items, travel bookings and event tickets.
How it works matters for safety. Google says Gemini uses an on-device virtual phone environment: it opens the app and interacts with its screen, like adding an item to your cart. If a task can be done through a connected app instead, Gemini uses that. You can watch live and stop the task. Taking over manually isn't available yet on Pixel 10 or Galaxy S26; on those you can only stop it. For the last step, such as confirming a booking or placing an order, you finish it yourself.
There's also a practical catch. If you ask Gemini to prepare an order and don't complete it, the items may stay in your cart.
Device assistance
The second power is simpler and more widely useful. Google's Gemini Intelligence page says device assistance lets Gemini manage alarms and timers, open apps, read and reply to notifications, control media and change settings. That's the "AI that controls your phone" most people actually use day to day.
What's coming for app developers
Google's developer post on making AI agents more helpful for Android apps describes two tracks. AppFunctions lets apps expose functions directly to agents, and Google compares it to MCP servers running on the device. Gemini already uses it for calendar, notes and tasks apps, and for Samsung Gallery on the Galaxy S26. The second track is UI automation for apps with no integration, which is what powers multi-step tasks.
The long-term picture: apps that opt in get fast, reliable function calls. Apps that don't get driven through their screens.
Samsung Galaxy AI
Samsung's Galaxy AI page leans on assistive features more than agents. Now Nudge reads what's on your screen and suggests relevant actions, like opening Calendar while you make plans in a text. Now Brief gives personalized briefings and now accepts your own prompts. Samsung also says Bixby can answer questions about the device "like a device AI agent" and search the web agentically.
For actually completing tasks across apps on a Galaxy phone, the current answer is Gemini. The Galaxy S26, Z Fold 8, Z Flip 8 and Tab S12 are all on Google's supported list above.
Open-Source Android AI Agents
If you want an agent that can try any app, not a curated list, you're in developer territory. All of these need a computer, a model API key and a phone with USB debugging or an emulator.
Mobilerun (formerly DroidRun)
The droidrun/droidrun GitHub repo now redirects to Mobilerun, and the README describes it as an open-source framework for mobile app testing, QA and running AI agents on real Android and iOS devices. It's MIT-licensed.
Setup installs a Portal app on the phone and enables its accessibility service. From there the agent combines accessibility trees with screenshots, so it can tap, swipe and type through multi-step tasks. It works with OpenAI, Anthropic, Gemini, xAI, Ollama, DeepSeek, OpenRouter and other OpenAI-compatible models. There's a CLI, a Python API, and a hosted Mobilerun Cloud if you don't want to run the agent locally.
Best for: app testing and QA, and developers building their own phone agents.
mobile-use
mobile-use from Minitap is an Apache-2.0 agent that controls an Android or iOS device from natural language. It can also scrape data out of apps into structured formats like JSON. The project says it was the first to complete 100% of the AndroidWorld benchmark. Its README also notes a real limit: it works poorly with games, because they don't expose accessibility data.
Best for: experimenting with end-to-end task automation and data extraction.
Mobile MCP
If you already use Claude Code, Codex, Cursor or Gemini CLI, Mobile MCP is the shortest path. It's an Apache-2.0 MCP server that lets any MCP client drive iOS and Android emulators, simulators and real devices. It reads the native accessibility tree first, which avoids image tokens, and falls back to screenshots and coordinate taps when it has to.
Best for: letting your existing coding agent test or operate a phone.
AndroidWorld, the benchmark behind the claims
When a project says it "beats" others, the test is usually AndroidWorld from Google Research. It runs on a live Android emulator with 116 hand-crafted tasks across 20 real apps, with randomized parameters so agents can't just memorize answers. Treat benchmark scores as a signal, not proof. A task suite on a clean emulator is easier than your real phone with notifications, logins and pop-ups.
Remote control over adb
Under most of these tools sits Android Debug Bridge, the command-line tool that talks to a device for installs, shell commands and input events. adb plus a model that reads screenshots is enough to build a basic agent yourself. The Gemini API's computer use tool even has a mobile environment for this. You still write the code that executes each action on the device. For how screen-reading agents work in general, see our explainer on the computer-using agent.
Safety: Accessibility Services, Banking Apps and Permissions
An agent that can tap anything on your screen can also tap the wrong thing. Here's what to know before you hand one your phone.
Accessibility access is powerful by design. Google's help page on restricted settings explains that an app with accessibility access can read content on your screen and interact with apps on your behalf. On Android 13 and up, some sensitive settings are restricted for certain installed apps until you explicitly allow them, and Google doesn't recommend allowing them unless you trust the developer. Scam apps often push people to turn these on.
Google Play doesn't allow autonomous agents on the Accessibility API. The Play policy on the AccessibilityService API says any use of it that lets an app autonomously initiate, plan and execute actions or decisions is prohibited. Rule-based automation ("if X happens, do Y") is still allowed, and verified tools for people with disabilities are exempt. That's why the open-source frameworks above install their helper apps outside the store, and why Gemini's feature runs on Google's platform-level UI automation instead.
Keep sensitive apps out of the loop. Google's own guidance for Gemini is to avoid screen automation for emergencies or sensitive information, and never to type sign-in or payment details into the chat. Take control and enter them in the app yourself. Apply the same rule to every agent here, and be stricter with open-source ones:
| Risk | What to do |
|---|---|
| Agent reads your banking or authenticator app | Use a separate test phone or emulator for experiments |
| Agent places an order or sends a message you didn't mean | Use tools that stop for confirmation, and check the final step yourself |
| A webpage or message contains hidden instructions | Don't let agents read untrusted content and act on it in the same run |
| Leftover access after testing | Turn off USB debugging and the helper app's accessibility service when done |
What Works Today vs. What's Still a Demo
| Works reliably now | Works, with caveats | Mostly demos |
|---|---|---|
| Setting alarms, timers and settings by voice | Gemini ordering rides, food and groceries in supported apps | "Plan my whole trip and book everything" across many apps |
| Replying to notifications | Open-source agents running scripted QA flows | Agents handling logins, 2FA and payments on their own |
| Summarizing what's on screen | Coding agents driving an emulator over MCP | Agents that work in any app, first try, on any phone |
The pattern is consistent. Agents do well on short, well-defined tasks in apps they've been tuned for. They struggle with long chains, unexpected pop-ups and anything that needs a password. If you want the bigger picture of where phones and desktops are heading, our explainer on the AI operating system covers the platform side.
Where Taku Fits
Android agents are improving fast, but today they cover a short list of apps and still need you to finish the last step. Taku doesn't replace Gemini or any of the frameworks on this list. If you'd rather run AI tools someone has already built on your Mac than wire up your own, that's where Taku comes in.
Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.
Key Points
- Gemini's multi-step tasks are the leading mainstream Android agent for third-party apps, in beta, on select devices, countries and apps.
- Gemini device assistance covers settings, alarms, notifications and media on a much wider range of phones.
- Galaxy AI is mostly assistive. Cross-app tasks on Samsung's newest phones run through Gemini.
- Mobilerun (formerly DroidRun), mobile-use and Mobile MCP are open-source developer tools that drive a phone over adb or an accessibility service.
- Google Play bans autonomous agents on the Accessibility API, so keep experiments on a test device and away from banking apps.
FAQ
Is there an AI agent that can use my Android apps for me?
Yes, if you have a supported phone. Gemini's multi-step tasks can book rides and order food or groceries in select apps on Pixel 10 and 11, Galaxy S26, Z Fold 8, Z Flip 8 and Tab S12 devices in a handful of countries. Newer devices add shopping, travel and tickets.
Can Gemini control my whole phone?
No. Multi-step tasks run apps in a secure virtual window and can only reach a limited set of apps, not the rest of your device. Separately, device assistance can change settings, open apps and handle notifications.
What happened to DroidRun?
The project now goes by Mobilerun. The old droidrun/droidrun GitHub repo redirects to droidrun/mobilerun, it's still MIT-licensed, and it adds a hosted Mobilerun Cloud alongside the local framework.
Is there an open-source Android AI agent I can run myself?
Yes. Mobilerun, mobile-use and Mobile MCP are all open source. Each needs a computer, a model API key, and a phone with USB debugging enabled or an Android emulator.
Is it safe to give an AI agent accessibility access?
Only if you trust the developer. Accessibility access lets an app read your screen and act in other apps. Test on an emulator or spare phone first, keep banking and authenticator apps out of it, and turn the access off when you're done.
Why isn't there an AI agent app on Google Play that uses all my apps?
Google Play policy prohibits apps from using the Accessibility API to autonomously plan and execute actions, except verified tools for people with disabilities. Agent features on Android come from Google's own system-level automation or from developer tools installed outside the store.