Give Claude Code an Android phone: Android MCP server setup.
Key moments
- 0:00 How do I give Claude Code an Android phone?
- 0:22 What is an Android MCP server?
- 0:49 How do I connect Claude Code to it?
- 1:20 Can Claude Code install my own APK?
- 1:40 How does the agent read the screen?
- 1:57 Watch it tap through a real app
- 2:16 Can a person take over?
- 2:28 What happens when it's done?
- 2:45 Does it work with Codex and Cursor?
- 2:59 Get started
Transcript
0:00 To give Claude Code an Android phone, add Phonebox's hosted Android MCP server with one command.
0:06 Claude Code then gets a real cloud Android phone it can see and control.
0:11 It reads the screen, taps, types, and installs apps.
0:14 No emulator, and nothing to install.
0:17 Here's the whole setup, and a real run, in about five minutes.
0:22 So, what is an Android MCP server?
0:24 MCP, the Model Context Protocol, is how Claude Code talks to tools.
0:28 Most Android MCP servers wrap ADB, which means you need an emulator, or a USB phone, on the same machine as your agent.
0:36 Phonebox is hosted instead.
0:38 The phone runs in the cloud, so it works from your laptop, a CI runner, or a cloud sandbox.
0:44 And the phone keeps its apps and sign-ins between sessions.
0:49 Connecting takes two steps.
0:51 First, get an API key.
0:52 The quickest way is to paste the setup prompt into Claude Code.
0:57 It signs you up from the terminal, and saves a key on your machine.
1:02 Then add the server: claude mcp add, with phonebox.dev/mcp and your key as a bearer header.
1:10 Now Claude Code has thirteen phone tools: create phone, observe, act, install app, live view, park phone, and more.
1:20 Now, a real run.
1:21 I ask Claude Code to install our sample app, Dockside Supply, and reorder last week's stock.
1:27 To install an APK, it calls create app upload, sends the file with one curl command, completes the upload, and calls install app.
1:36 About three seconds later, the app is on the phone.
1:40 Then it looks.
1:41 The observe tool returns the screen as a short, numbered list: every button and text field, with its label, plus a screenshot.
1:49 That's far less to read than raw pixels, and the agent can tap by label, instead of guessing coordinates.
1:57 Now it acts.
1:58 Tap Orders.
1:59 Tap Reorder.
2:00 Set oat milk to six crates.
2:02 Place order.
2:02 After each step, it reads the screen again to check the result.
2:07 And Phonebox never repeats a write on its own, so a slow response can't turn into a double order.
2:14 Order placed.
2:16 Some steps need a human, like a one-time sign-in code, or a payment.
2:20 The live view tool creates a link you open in a browser, to watch and control the phone, and then hand it back to the agent.
2:28 When it's done, Claude Code parks the phone.
2:31 A parked phone costs nothing, and keeps its apps, files, and signed-in accounts, so the next task starts where this one ended.
2:39 And if the agent forgets, the phone parks itself after five idle minutes.
2:45 The same server works with Cursor, Codex, or any MCP client that can send an authorization header.
2:52 And if you'd rather skip MCP, there's a CLI, a TypeScript SDK, and a REST API.
2:59 That's how to give Claude Code an Android phone: one MCP server, and real cloud phones.
3:05 The setup guide and the full transcript are at phonebox.dev.
Try it yourself
Paste Read https://phonebox.dev/setup.md and set me up with Phonebox into Claude Code, Codex or Cursor. Then read give your coding agent an Android device, the Android MCP server guide, or cloud phones for AI agents.
Watch on YouTube.
The MCP tool calls and their results come from a real run against phonebox.dev/mcp, and the phone footage is that phone's own screen recording. The Claude Code window is a recreation that shows those calls. Dockside Supply is a sample app built for demos.