Description
You'd like an AI to run a routine on your phone — scroll through a feed, tidy up a list, type a note in Chinese — but the apps on your phone are out of its reach, so you end up tapping through it yourself.
iPhone-use plugs Codex straight into your real iPhone. Connect a USB cable and say something like "open Notes, start a draft and type this", and Codex opens the app, reads the page, taps, swipes and types. A sidebar shows an iPhone with a live screen, with a cursor hint on every tap and drag, so you can see exactly where it is.
No jailbreak, and no separate Appium server — the phone side runs WebDriverAgent. Installation is handed to Codex too: paste the prompt from the repo and it checks Xcode, Python, Node.js and the phone connection, then walks you through signing and the first launch.
Reads the page, falls back to screenshots: it prefers a compact control tree, and when a target can't be found or is covered, it switches to a screenshot and converts pixels to tap coordinates instead of tapping blindly.
Chinese and long-text input: full Unicode text in one go, long text typed in chunks with resume after an interruption. Typing and submitting are separate — nothing is submitted by default until the model has checked.
Batched actions: known taps, typing and waits are bundled into one batch, cutting a lot of model round trips.
Scroll-to-find and list collection: bounded scrolling to find a target, de-duplicated reading, and a report of how far it covered and what is still unconfirmed.
Live phone screen: an iPhone frame with device status and refresh, home and screenshot buttons; it tells a locked screen from a deliberate pause, and one click reconnects a dropped preview.
Careful failure handling: a timed-out action may already have landed, so it reads the real state before deciding and never blindly replays taps or submits; multiple processes share one operation lock so they don't collide.
Data stays on your machine: everything goes through a local USB forward, signing material and device data never leave your computer, and the screen stream isn't written to disk; passwords, codes and Face ID are left for you to complete.
iPhone-use plugs Codex straight into your real iPhone. Connect a USB cable and say something like "open Notes, start a draft and type this", and Codex opens the app, reads the page, taps, swipes and types. A sidebar shows an iPhone with a live screen, with a cursor hint on every tap and drag, so you can see exactly where it is.
No jailbreak, and no separate Appium server — the phone side runs WebDriverAgent. Installation is handed to Codex too: paste the prompt from the repo and it checks Xcode, Python, Node.js and the phone connection, then walks you through signing and the first launch.
Features
Reads the page, falls back to screenshots: it prefers a compact control tree, and when a target can't be found or is covered, it switches to a screenshot and converts pixels to tap coordinates instead of tapping blindly.
Chinese and long-text input: full Unicode text in one go, long text typed in chunks with resume after an interruption. Typing and submitting are separate — nothing is submitted by default until the model has checked.
Batched actions: known taps, typing and waits are bundled into one batch, cutting a lot of model round trips.
Scroll-to-find and list collection: bounded scrolling to find a target, de-duplicated reading, and a report of how far it covered and what is still unconfirmed.
Live phone screen: an iPhone frame with device status and refresh, home and screenshot buttons; it tells a locked screen from a deliberate pause, and one click reconnects a dropped preview.
Careful failure handling: a timed-out action may already have landed, so it reads the real state before deciding and never blindly replays taps or submits; multiple processes share one operation lock so they don't collide.
Data stays on your machine: everything goes through a local USB forward, signing material and device data never leave your computer, and the screen stream isn't written to disk; passwords, codes and Face ID are left for you to complete.
