Playbooks · Coding
iPhone-use: a phone agent, no Mac needed
Written November 10, 2026
iPhone-use is an open-source iPhone app where a language model controls the phone it runs on. A Mac is needed once, to make the pairing file. After that, the phone works alone.
The problem
Phone agents usually need a computer in the loop: the model runs on a Mac, the phone stays plugged in, and the Mac drives it. That works at a desk and nowhere else. iPhone-use moves the agent onto the phone itself, so it can run a task while the phone is in your pocket.
How it works
From the iPhone-use repository: the app mounts a developer image, then uses a loopback VPN to screenshot and control the iPhone it runs on. Each step, the screenshot goes to the model, which answers with a tool call — tap, long press, swipe, type, or open app — until the task is done. A Live Activity shows progress with a Stop button, and every run is saved as a chat. You can start a run from the chat, from Siri, from a "Run Agent" Shortcut, or from an iphoneuse://run?prompt=… link.
The demo on the project page has the agent pick a Lyft ride to the Eiffel Tower without booking it: it leaves its own app, opens Lyft, types the destination, picks it, and stops at the ride options.
What you need
- iPhone on iOS 26 or later, with Developer Mode on. Both are hard requirements.
- A Mac, once. Plug the phone in over USB and make a pairing file with the
idevicetools. Pairing files expire, so make a new one when the app says so. - A vision model. An OpenRouter key, or any OpenAI-compatible endpoint whose model takes images and tool calls.
- A way to install it. With a paid Apple team you build from source, which gives the app its own built-in VPN. With a free Apple ID you build with
BUILTIN_VPN=NOand install the LocalDevVPN app alongside it, and re-sign every 7 days. It needs Developer Mode and developer services, so it cannot go on the App Store.
What the project measured
The project runs its own benchmark: a script replays 11 recorded steps (open Lyft, search a destination, pick a ride, stop before paying; write a note) against each model through OpenRouter, twice, for 22 decisions per model. The numbers below are the project's, from November 2026:
| Model | Correct | Median | Cost per decision |
|---|---|---|---|
| Claude Haiku 5.5 (default) | 91% | 1.3 s | $0.0003 |
| Claude Sonnet 5.5 | 91% | 1.2 s | $0.0055 |
| Claude Opus 5 | 95% | 2.1 s | $0.014 |
| Claude Opus 5, fast tier, reasoning off | 100% | 1.4 s | $0.051 |
| Claude Opus 5.5 | 86% | 1.8 s | $0.011 |
That is 11 steps, not a broad test. The cheapest model gets most steps right for a fraction of a cent each; the most expensive gets them all right for about five cents a step.
Limits
- Expect a few seconds per step, most of it the model thinking.
- It needs a loopback VPN, and iOS allows one VPN at a time, so Tailscale and similar stop while it is on. On cellular data, iOS refuses new loopback VPN connections, so the app runs a shortcut that turns data off while it connects, then back on.
- iOS shows a "Viewed remotely from Mac" notice on every run when its developer services are used, even though the "Mac" here is the app itself.
- The app asks for Developer Mode, a VPN configuration, Local Network access, and Location set to "Always" to keep runs alive in the background. The microphone is only for voice prompts.
- Nothing here was tested by anyone outside the project. The benchmark above is the project's own, on its own recorded steps, and the pairing, install, and VPN steps are the project's instructions, not a second person's walkthrough.
- Only automate devices and accounts you own or may use, and respect the terms of the apps involved. The agent acts with your hands: watch it, and keep Stop within reach.
Related playbooks
- T3 Code: drive your agents from your phone
- Stop a sleeping Mac draining your Claude Code limit
- Run an open-weight agent with egress denied: what the Qwen sandbox incident does and does not prove
- Addy Osmani's 25 agent skills for coding agents
Sources
Back to: Coding playbooks hub · Also: Recalc: agents inside spreadsheet cells