No jailbreaking, no iPhone app installation: Open-source project Poki lets AI directly operate iPhones.

Beating AI Express News: Developer Joseph Zhou has open-sourced Poki, a tool that allows AI agents like Codex and Claude Code to directly control real iPhones. Existing iPhone automation projects typically either use XCTest/WDA to install test applications or rely on external hardware to simulate operations. Poki takes a different approach: it directly calls Apple’s device developer services, reads the iPhone’s native accessibility control tree to retrieve buttons, text, and positions, then sends commands including taps, swipes, inputs, and even pinch-to-zoom actions to the device. This eliminates the need for AI to guess coordinates via screenshots at every step, or install test apps on the iPhone. After initial USB pairing, Poki supports wireless operation on the same local network. Beyond integrating with Codex and Claude Code via MCP, Poki also includes a Mac-based Agent that can receive tasks via terminal or iMessage. It can store successfully executed and verified operation workflows, reusing them for identical tasks in the future—some tasks do not require additional model calls. However, the Mac must remain powered on for Poki to function. Real-device operations are currently slow: the author tested on an iPhone running iOS 27.3, with 12 consecutive taps on a search box in a chat app all working, each taking around 8 seconds; filling an input box takes ~15 seconds, with some buttons still not responding properly. Poki’s strengths include direct system control reading, wireless operation, and multi-touch support, but it requires full Xcode, developer mode, and UI Automation. Its underlying layer relies on Apple’s developer APIs, meaning future system updates may impact compatibility.