Image: GitHub
Why it mattersA team can give a coding agent a real iPhone to drive. The agent taps, types and swipes while a person watches, and can be taken over at any moment.
A software engineer named Aleksey Smolenchuk has released a new open-source tool. It lets an AI coding agent run tasks on a real iPhone. The project is called phoneuse. The repo was created on 7 November. One week later, on 14 November, it had 657 stars and 61 forks on GitHub. The author's GitHub bio lists prior work at OpenAI, Netflix, Opendoor and Uber.
phoneuse does two things. It mirrors the phone's screen back to a Linux PC, Mac or Windows PC over AirPlay. It sends mouse clicks, drags, scrolls and keyboard presses to the phone as Bluetooth input. Both sides sit on the same Wi-Fi network. The phone is not jailbroken and does not have to be on a cable. The agent sees the screen as a window and taps on it like a user.
What the agent can do
A connected agent like Claude Code can tap, type, swipe by dragging, and scroll with the mouse wheel. The author lists these four actions in the README. A permission prompt on the iPhone can ask the user before each action. The user can stop the agent and take the mouse back at any moment.
The companion iPhone app sits on the App Store under listing 6832337610. Linux users get a Flatpak build. Mac users get a signed DMG for macOS 15 and later. Windows users build the tool from source. The whole project is MIT licensed.
The optional hardware bridge
The Mac needs no extra parts. The Linux and Windows builds use an optional hardware bridge to send Bluetooth. The bridge is a LilyGO T-Display board that joins the same Wi-Fi as the iPhone and the computer. It then presents itself to the phone as a Bluetooth keyboard and mouse. That keeps the computer's own Bluetooth free for a headset or a keyboard.
What state the project is in
Smolenchuk marks the project "alpha software" at the top of the README. He asks for bug reports. The repository shows 371 commits on the master branch, with the last push on 13 November. There are 10 open issues. The reach today is one phone per person. One person must stay near the computer to catch a bad step.
A reader looking for a reason to try it this week has two real ones. A QA team can give an agent a real device and a login, then read the recording at the end. A support team can let an agent walk a user through a fix, with the person watching. Both jobs used to need a cable, a build of the app, and Apple's own test tools. phoneuse needs only Wi-Fi and the app on the phone.
The reader looking for a reason to wait has one. Driving a real phone with no sandbox puts a lot of trust in one script. A missed step is a real tap on a real screen, in a real account. The permission prompt and the take-over button are the safety net. A team that wants more than that will have to add it themselves.
Source
github.com/lxe/phoneuse on GitHub. Repository metadata from api.github.com/repos/lxe/phoneuse. Author profile at github.com/lxe.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.