A cloud phone for AI agents is a phone that runs in a data center and that an agent controls over an API or MCP: it reads the screen as text, taps, types and installs apps, and code starts, stops or deletes it. It gives an agent the apps that have no website, and gives a coding agent a place to run the app it built without a device on your desk.
Sources reviewed on November 9, 2026: Mobilerun pricing, Firebase Test Lab pricing, GeeLark and mobile-mcp. Mobilebox facts come from the API reference.
Three things are called a cloud phone
| Kind | Built for | How it is driven | Example |
|---|---|---|---|
| Agent cloud phone | An agent that acts in apps, step by step | REST or a hosted MCP server: read the screen, act by number | Mobilebox, Mobilerun, phonebox.dev |
| Device farm | Test suites across many devices | A test framework, run by run | Firebase Test Lab: $1 an hour virtual, $5 physical |
| Multi-account phone | People running many app accounts | A dashboard and scripts | GeeLark: "Your AI copilot for social media marketing" |
Only the first kind is made for an agent that decides its next step from what it sees. A device farm runs a test you wrote in advance and returns the result. A multi-account phone is made for a person at a dashboard.
What the agent actually does
An agent cloud phone turns a phone into a loop of tool calls. On Mobilebox:
create_phone → observe → act → observe → park_phoneobserve returns the screen as numbered elements, one per line, such as [3] Text "Search" (540,288). act takes a list of taps, typing and swipes by number and returns the new screen. Your agent’s own model chooses each step; the phone runs no model of its own. The agent guide walks through the loop, and the live view lets you watch it.
Cloud phone, local emulator or your own phone?
- A local emulator with mobile-mcp: free and fast on a laptop with hardware acceleration, but it lives on that laptop, and it stalls in cloud agent sandboxes and CI runners without KVM. See cloud phone or local emulator.
- Your own phone over USB: real hardware and your own accounts, one phone, on your desk.
- A cloud phone: reachable from anywhere an agent runs, one per task if you want, and kept between runs when you park it.
How to choose one
- Where does your agent run? A cloud agent or a CI job needs a URL and an API key, not a USB cable.
- Android, iPhone or a real device? Apps that check for real hardware need a physical phone; see real devices or cloud phones.
- Does it keep logins? Mobilebox keeps apps, data and logins when you park a phone. Mobilerun says "Stopped phones keep apps and data and bill no minutes."
- What do idle and boot time cost? Mobilebox bills $0.01 per ready phone-minute, per second, with boot time and parked phones free. Mobilerun charges $0.03 per extra minute on its paid plans.
- How fast does it start? We measured from under a second to 23 seconds from create to first screenshot across three services; see the benchmarks.
Six agent cloud phones side by side: cloud Android phones for AI agents, compared. Which one fits which job, with measured start times: the best cloud phones for AI agents in 2026. Mobilebox calls its phones mobile sandboxes: a cloud phone your agent can use and throw away.