How to give an AI agent an iPhone with MCP
Disclosure: I'm the maker of thumbs, a remote MCP server for this. The first half applies to any approach. For a web-app-specific walkthrough, see how to test your web app on an iPhone with an AI agent.
Your agent can browse the web, call APIs and edit code. But much of what people do happens in iPhone apps, and many services only answer in an app. To use those, an agent needs a phone.
How it works under the hood
- Something that can operate the phone. On iOS that's usually WebDriverAgent, which drives the UI through the same XCTest machinery Apple's own UI tests use. It can take screenshots, read the accessibility tree, tap, type and swipe.
- A translation the model can use. A screenshot alone leaves the model guessing pixel positions. Better: a screenshot plus a numbered list of every tappable element with its label and position, so the model says "tap 7".
- A protocol the agent speaks. MCP (Model Context Protocol) is how Claude Code, the Claude app and other agents discover and call tools. A "remote" MCP server lives at a URL and uses OAuth sign-in, so there's nothing to install.
Local or remote, Simulator or real device
| Local MCP server | Remote MCP server | |
|---|---|---|
| Setup | Mac, Xcode, config, signing for real devices | Add one URL, sign in |
| Phone | Yours, on your desk | Someone else's, over the internet |
| Best for | Full control, offline | No Mac, no cable, shared tooling |
| iOS Simulator | Real iPhone | |
|---|---|---|
| Apps | Safari and Apple's built-in apps | Apps you install (banking and Apple Pay apps still don't work) |
| Logins | Shared, so none | Yours, kept between sessions |
| Fidelity | Close to Safari, not hardware | The real thing |
thumbs is a remote server. Today its free early access runs on the iOS Simulator, shared between testers. A real iPhone for one customer is planned; it isn't available yet and has no date or price.
Connect it
Claude Code
claude mcp add --transport http thumbs https://mcp.usethumbs.com/mcp
Then /mcp and sign in with Google. Finish the sign-in in a browser on the same computer while Claude Code waits.
Claude app: add the same URL as a custom connector.
Other MCP clients: any client that supports remote servers over HTTP with OAuth sign-in connects the same way.
Only accounts let in from the waitlist can use a phone.
What agents are good at on a phone
Reported by the maker, measured on a real iPhone 17 Pro on 23 Oct 2026 with Claude Code (not on the Simulator that early access runs on today): - Reading a setting (iOS version): 7 turns, 31 s. - A Safari search and reading back the first result: 8 turns, 27 s. - Walking a language app's 19-screen onboarding to its paywall, declining every prompt: 3.2 minutes. - 296 commands in 20 minutes with no failures; a screenshot in 63 ms.
Taps take 0.4 to 0.8 s, so don't expect a game controller.
Limits and safety
- Banking apps and Apple Pay don't work, and nothing behind Face ID.
- Use test accounts. In early access the Simulator is shared between testers.
- Decide in advance what the agent must not do (purchases, sending messages) and say it in the prompt.
FAQ
Can Claude control an iPhone? Yes, through an MCP server that exposes phone actions to it.
What is an MCP server for iPhone? A service that exposes "see the screen, tap, type, swipe, open app" as tools an agent can call.
Do I need to jailbreak? No. WebDriverAgent runs as a normal XCTest bundle on a developer-enabled phone.
Can it use my own apps and accounts? That's what the planned own-iPhone option is for; it isn't available yet and has no date or price. Not on the shared Simulator.
How do I get access? Join the waitlist; people are let in one at a time. Early access is free.