Skip to content

No jailbreak required, no app installation on iPhone: open-source project Poki lets AI directly operate your iPhone.

Nov 6, 16:48

Beating AI News Flash: Developer Joseph Zhou has open-sourced Poki, enabling AI Agents such as Codex and Claude Code to directly control a real iPhone.

There are already quite a few iPhone Use projects—some install test programs via XCTest/WDA, while others rely on external hardware to simulate operations.

Poki takes a different route: it directly calls Apple's device developer services, reads the iPhone's native accessibility control tree to obtain buttons, text, and positions, and then sends commands to the phone such as taps, swipes, text input, and even two-finger pinch-to-zoom.

This way, the AI does not have to guess coordinates from screenshots at every step, nor does it need to install test programs on the iPhone. After the initial pairing via USB, it can also operate wirelessly within the same local network.

In addition to connecting to Codex and Claude Code via MCP, Poki also comes with its own Agent running on a Mac, which can receive tasks through the terminal or iMessage. It can remember operation workflows that have been successfully executed and verified, and directly reuse them the next time the same task comes up—some tasks do not even require calling the model again. However, the Mac still needs to stay running.

Real-device operation is currently still relatively slow. The author tested on an iPhone running iOS 27.3: for a search box in a chat app, 12 consecutive text taps all took effect, but each took about 8 seconds; filling in an input field took about 15 seconds, and some buttons still could not be tapped properly.

Poki's advantages are that it can directly read system controls and supports wireless operation and multi-touch, but it requires the full Xcode, Developer Mode, and UI Automation. It relies on Apple's developer interfaces at the underlying level, and future system upgrades may also affect compatibility.

Source