Cloud phones for AI agents: what they are and how to choose one.

Watch the 2½-minute explainer: what cloud phones for AI agents are, how they compare with emulators, device farms and phone numbers, and how an agent drives one.

Why agents need a phone

Agents got good at the web and the terminal. Then they hit the apps. A huge part of real work lives inside Android apps that have no public API and no web version: banking and delivery apps, internal field tools, messaging apps, and the mobile app you are building yourself.

A cloud phone closes that gap. The agent gets the same device a person would use, with a real app store, real sign-ins and real touch input, and it gets it on demand, from wherever the agent runs.

  • Test your own app. Install the APK your agent just built and have it walk the screens it changed.
  • Work in apps with no API. Read a status, fill a form or check a notification inside a mobile-only app.
  • Keep signed-in state. Sign in once; the phone keeps the account between sessions.
  • Give every customer a phone. Products built on agents can give each end user their own persistent device.

How an agent uses a cloud phone

The loop is the same in every interface: observe → act → verify → park → resume. Phonebox returns the screen as a short numbered list of what is on it, so a model can choose an element by its label instead of guessing pixels. Screenshots are there when the agent wants one.

What the agent reads
snapshot snp_4f2k7m3q6z3a
app com.android.launcher3
keyboard hidden
screen 1080x2400
[1] EditText "Search apps" #search clickable editable (540,250)
[2] TextView "Chrome" #icon clickable (200,1900)
[3] TextView "Settings" #icon clickable (500,1900)
[4] TextView "Play Store" #icon clickable (800,1900)

The agent taps "Search apps" by its text, types, and reads the screen again to confirm the result. When it finishes, it parks the phone: billing stops, and the apps, files and signed-in accounts stay for next time. How AI agents use mobile apps explains each step of the loop.

agent.ts
import { Phonebox } from "phonebox";

const pb = new Phonebox(); // reads PHONEBOX_API_KEY
const phone = await pb.phones.create({ name: "signup-test", idempotencyKey: `signup-${Date.now()}` });
try {
  if (phone.status !== "ready") await phone.waitUntil("ready");
  console.log(await phone.look());
  await phone.tap({ text: "Sign in" });
  await phone.waitFor({ text: "Welcome" }, { timeoutMs: 20_000 });
} finally {
  await phone.park(); // parked phones cost nothing and keep their apps
}

Three kinds of cloud phone

Search for “cloud phone” and you will find very different products under one name. Pick the one built for your job.

Cloud phone categories compared
KindBuilt forHow you control itTypical pricing
Agent-native cloud phonesAI agents operating Android appsMCP, API, CLI and SDK with a text view of the screenPer minute or per second of use
Multi-account cloud phonesRunning many social and marketplace accountsA dashboard, browser profiles, proxies, RPA scriptsMonthly plan per device
Device farmsQA teams running test suitesAppium, Espresso and test frameworksPer device-minute, or plans with parallel slots

Multi-account tools such as GeeLark and Multilogin's cloud phones are designed around profiles, proxies and social media management. Device farms such as BrowserStack, AWS Device Farm and Firebase Test Lab are designed around scripted test runs. Both can be wired to an agent, but neither was designed around an LLM reading a screen and deciding what to do next.

Phonebox is agent-native: the interfaces, the screen format, the billing and the safety rules all assume the driver is a model. For a vendor-by-vendor view, read the best cloud phones for AI agents.

Why Phonebox is the agent-first cloud phone

Most cloud phones were built for something else. Some are antidetect tools for running social media accounts from a dashboard; others are device farms for QA test suites. They added an API later. Phonebox started with the agent: every interface, default and error was designed for a model driving the phone, not a person clicking a console.

  • Your agent signs you up. One prompt in Claude Code, Codex or Cursor: the agent installs the CLI, opens a sign-in link for you, and saves its own key. No dashboard setup, no key pasted into the chat.
  • A screen a model can read. The screen comes back as a short numbered list of elements, so the agent taps “Sign in” by its label instead of guessing pixels. Screenshots are there when it wants one.
  • Errors that say what to do next. Every refusal carries a next step and whether it is safe to retry, so the agent recovers without a human reading logs.
  • Safe by default. Phonebox never repeats a write to a phone on its own; an unknown outcome tells the agent to look before it acts again. Creates take idempotency keys, so a retried request never makes a second phone.
  • Built for forgetful agents. Idle phones park themselves and cost nothing while parked, and every session's credit is reserved up front, so a balance can never go negative.
  • Every interface an agent uses. A hosted MCP server with tool hints, a CLI, a TypeScript SDK, REST, an agent skill and llms.txt.
  • A human when it matters. A live view link hands the phone to a person for a one-time code or a payment, then back to the agent.
  • Your agent, not ours. No bundled agent and no agent credits: Phonebox is the phone, billed per second, for whichever agent you already run.

What to look for

  1. Agent interfaces. A hosted MCP server and a CLI your coding agent can run, plus an SDK and REST for your product.
  2. A screen the model can read. Structured text with element labels beats screenshots alone: fewer tokens and fewer mis-taps.
  3. Persistence. Apps, files and sign-ins should survive between sessions, or every task starts with a login.
  4. Pay for use, not for idle. Per-second billing, an idle timeout and free parking keep a forgotten phone from billing all night.
  5. Human handoff. A link that lets a person finish a one-time code or a payment, then hand the phone back.
  6. Safe writes. An agent should not repeat a tap because a response was slow. Phonebox never retries a write to a phone automatically.
  7. Honest scope. Android device control is not telephony. If you need calls or SMS, you need a phone-number provider as well.

Phonebox at a glance

Phonebox facts, from the current documentation
Phonebox
DevicePersistent cloud Android phone
InterfacesHosted MCP, CLI, TypeScript SDK, REST, agent skill
ScreenNumbered text view of the screen, optional JPEG or PNG screenshot
ActionsTap, type, swipe and scroll, back, open apps, wait for elements
AppsInstall from an https link to an APK, upload your own APK up to 500 MiB, or use the app library
Device settingsClipboard, location, locale, timezone, screen recording
Human handoffLive view link
Price$0.06 per phone-minute, billed per second, one-minute minimum
Parked phonesFree, with apps, files and sign-ins kept
StartNew accounts start with a free 7-day trial, card required.

Give your agent a phone in two minutes

Paste this into Claude Code, Codex or Cursor. You do not need an account first; your agent signs you up from the terminal.

Paste into your agent
Read https://phonebox.dev/setup.md and set me up with Phonebox from here: get me signed in and ready, then find out what I want to use phones for and help me with that.

Next: give your coding agent an Android device, connect the Android MCP server, or test your Android app with an AI agent.

Questions

What is a cloud phone for AI agents?

A cloud phone for AI agents is a real Android device hosted in the cloud that software agents control through an API or MCP server. The agent reads the screen, taps, types, installs apps and moves files, with no phone on your desk and no local emulator. Phonebox is a cloud phone built for this: one API call starts an Android phone your agent can see and control.

Can Claude Code, Codex or Cursor control an Android phone?

Yes. With Phonebox, you paste one setup prompt into Claude Code, Codex or Cursor. The agent installs the Phonebox CLI, signs you in, and can then read an Android screen, tap, type and install APKs. MCP clients can also connect to the hosted server at https://phonebox.dev/mcp.

How much does a cloud phone for an AI agent cost?

Phonebox costs $0.06 per phone-minute, billed per second with a one-minute minimum per session. A parked phone costs nothing and keeps its apps and signed-in accounts. New accounts start with a free 7-day trial, card required. Multi-account cloud phones are usually sold as monthly plans per device instead.

Is a cloud phone the same as a virtual phone number?

No. A virtual phone number gives you calls and SMS. A cloud phone gives an agent an Android device: the screen, the apps and touch input. Phonebox is a cloud phone and does not include a phone number, voice calls or SMS.

Do I need an Android emulator on my machine?

No. The phone runs in Phonebox's cloud, so your agent can drive it from a laptop, a CI runner or a sandbox that cannot run an emulator. The agent only needs network access and an API key.

Can a human take over the phone?

Yes. A live view link opens the phone in a browser so a person can watch and control it, for example to enter a one-time code or approve a payment, and then hand it back to the agent.

What is the best cloud phone for AI agents?

For agents, Phonebox: it is the agent-first cloud phone. Your agent signs you up from the terminal, reads the screen as structured text, gets errors that say what to do next, and connects through a hosted MCP server, CLI, SDK or REST, at $0.06 per minute billed per second with free parked phones. Multi-account cloud phones such as GeeLark and Multilogin were built for social media dashboards, and device farms such as BrowserStack for QA test suites.

Does Phonebox support iPhone or iOS?

No. Phonebox provides Android devices only.

Get started with Phonebox. New accounts start with a free 7-day trial, card required.

The Phonebox team

← All guides