A $7 board that plugs into an iPhone and lets AI tap the screen for you: esp32-iphone-agent, 26 days, 4400 stars

Henry Yuan, a developer in New York, plugged a fingernail-sized ESP32 dev board into a stock iPhone so AI can watch the screen and tap for itself. No Mac, no developer account, no jailbreak. The repo gained 4436 stars and 596 forks in 26 days. It took off for three reasons: perfect timing, a pitch you can explain in one sentence, and both the hardware and AI circles were willing to share it. Below: what it is, where the stars came from, who it's competing with, and where to start if you want to build something that earns stars the same way.

One alarm-clock video

On Oct 12, 2026, Henry Yuan (on X as @iamhenryyuan, GitHub account registered in 2016, and this is his only public repo) posted on X in both Chinese and English: "A $7 ESP32 lets an AI agent control a stock iPhone. No Mac, no jailbreak…" In the attached video, a Seeed XIAO ESP32-S3 is plugged into the iPhone's USB port. He gives the AI one instruction: "Set a repeating alarm for 8 a.m. on Mondays and Tuesdays, then save." The AI exits the app on its own, opens Clock, spins the wheel, ticks the days, and saves, 22 steps in all, with no one touching the phone.

The English post got 210k views and 242 likes; the Chinese version got 213 likes (on X, as of 11-06). The next day, the major X account @RoundtableSpace reposted it with the line "Apple never built this, so someone hacked it with $7", 92k views. In Oct 14, Hackaday published "AI Controls iPhone Via ESP32", and the repo gained 1072 stars that day, its single-day high so far.

What it actually is

Broken down, the method isn't complicated. It uses features built into iOS that are open to everyone:

• Eyes: the board runs its own Wi-Fi hotspot, which shows up in the iPhone's screen-mirroring list as *ESP32 Agent*. The phone projects its screen to it, and the board doesn't decode anything. It passes the H.264 video stream straight through to the app on the phone.
• Hands: to the iPhone, the board is just an ordinary USB mouse and keyboard. With AssistiveTouch turned on, a mouse click equals a finger tap, and a drag equals a swipe.
• Brain: the app sends the screenshot to any model with an OpenAI-compatible API that can read images. The model decides where to tap, swipe, or type next.

The earliest version still needed a computer running an AirPlay receiver. In Oct 15, he moved that step into the board as well. The README says "The computer is gone," and at runtime all you need is a phone and a board. This update post got 807 likes, the most of any of his posts. Later came a command line and an MCP server (npx -y esp32agent mcp), so Claude Code and Cursor can connect directly and operate an iPhone. In Nov 3 the app went to TestFlight, and in Nov 6 it connected to DeepSeek Harness.

Its weak points are listed in the README itself: the screen is only 480×1040 at up to 5 frames per second, each step takes several seconds, and most of the time goes to waiting on the model. The pointer is relative, so it needs calibration first. Landscape isn't supported. The lock screen, password entry, and Apple Pay can't be mirrored because iOS blocks it, so the AI can't see them. While it runs, the board occupies the phone's Wi-Fi, so only cellular data works. And "no Mac" applies only at runtime: the first flash of the board still requires a computer.

Why this one, and why this month

Demand peaked that very week. In early October, "iPhone-use" suddenly took off in Chinese X circles. zhongerxin/iPhone-use was created in Oct 6 and got 4909 stars in a month. The same week, actell (10-09) and phone-agent (10-06) also appeared. Everyone was asking how to let an agent operate an iPhone, and this repo was created in Oct 11, just in time.

It bypassed the gate everyone else was stuck on. Similar solutions nearly all required a Mac, Xcode, developer signing, or enabling developer mode (listed one by one in the table below). Its title spelled it out: "$7 / No Mac / no developer account / no jailbreak." Even an outsider understands it at a glance, and they're willing to share it.

The same idea had been tried before without taking off. zhaodashuaibi/VersaiOS already used an ESP32 Bluetooth mouse plus an AirPlay receiver on a computer to control an iPhone back in March, seven months earlier. It has 9 stars now. boboabw/phone-agent came out the same week and also used an ESP32 with AssistiveTouch, but it skipped the screen-viewing step and has 21 stars. The extra step this one took was porting the AirPlay receiver onto a $7 board. That made "zero dependencies" a credible claim, and the engineering feat itself became a reason for others to share it.

Two crowds shared it at once. The AI agent crowd shared it on X, and the maker crowd shared it on Hackaday, Elektor, Hackster, and r/esp32. A pure-software iPhone-use project wouldn't make it onto Hackaday, but this one did, thanks to its physical photos. The board is available on Amazon for $7, so viewers felt "I could try this too." The fork-to-star ratio is about 13%, which is high for a hardware project and suggests many people actually tried it. How many actually got it running is not visible from the outside.

Where the stars came from

The growth pattern is typical: 10-17 was 2802, then it grew 80 to 170 a day before slowing down; by early November it was only adding thirty or forty a day. After 11-03 hit TestFlight and 11-06 posted the DeepSeek thread, it climbed back to 60 to 70 a day (GitHub daily stars, 11-06). Over the last 7 days: +343; the 7 days before that: +548. The overall trend is slowing, and it depends on posting a new thread with every release to push it up in waves.

Who it's competing with

Copycats have already arrived: ananhq/anan-iphone-use (10-13, ESP32 MCP firmware, 150 stars), ghosthand-dev/ghosthand (10-27, USB plus Bluetooth, 37 stars), and JosssphZhou/poki (11-05, pure-software MCP; its pitch directly contrasts itself with "no $7 board," and it has between a hundred and three hundred stars). There are also people riding the hype and poisoning it, pasting "one-click install" commands into issues, and the README now has a warning at the top about it. Apple's own Siri and Apple Intelligence still don't allow third-party models to operate arbitrary apps, and that is the precondition for this whole category of project to exist. The day Apple opens that up, this barrier disappears.

The author's move toward a product is also evident: they've set up a brand, Handsy, and a website, handsy.tech. The README links to Amazon affiliate links for buying the board and cables, the app is waiting on App Store review, and the first line of the README asks anyone who builds on it to credit the author.

If it were you

In one line: B. This wave of "AI operates iPhone" already has two projects with four to five thousand stars; the ones that followed copied it and only got dozens to hundreds. But the recipe of "a cheap board that acts as the AI's eyes and hands" can be reused once more on another device.

Who it's for: developers who can write ESP32 or other microcontroller firmware, understand USB HID and video streams a bit, and ideally already have some presence on X or in hardware communities, and who can publish in both Chinese and English.

Who it's not for: people who want to build another "AI controls iPhone." The original author, iPhone-use, and poki have already taken the "hardware," "software," and "no hardware" positions, so a new one won't have anything new to say in its title. It's also not for people who only write software and don't want to touch hardware. This category broke out in hardware circles precisely because of that physical board.

How to do it: swap in a device that is closed, has no automation interface, but can be screen-mirrored and accepts USB keyboard and mouse input, and copy the recipe over as-is. Candidates: professional apps on iPad, game consoles like Switch or PS5 (HDMI capture plus USB controller emulation), TV boxes, car head units, and e-readers like Kindle (inferred; feasibility not verified for each). Follow its title format: "A $X board that lets AI operate ___, no Y, no Z." The second path is to fix what it did poorly, like cutting each step from several seconds to under one, supporting landscape mode, or shipping a ready-made task library (auto check-in, auto form filling). But stars on that path mostly flow back to the original project, so it's hard to break out on your own.

How to get users: follow its playbook. The first post needs a full task video showing the AI completing a specific job from start to finish. Publish in Chinese and English on the same day. Submit to Hackaday the same week, and post project pages on Hackster, Instructables, and r/esp32. Then pair every release with a new video to stretch one burst into several waves. Put a comparison table in the README against similar projects, spelling out "they need X, I don't."

When to move: the topic of "AI agents operating real devices" is hot right now, but new projects arrive every week. Ideally you ship a first version you can record video of before late December. The later it gets, the less novel "yet another thing that lets AI control ___" feels to readers.

Signs you've got it wrong: (1) In November two or three more "AI controls a certain device" hardware projects shipped and none passed 500 stars, meaning people only care about the iPhone and switching devices won't work. (2) Apple or Google opens system-level third-party agent operation of apps, which removes the reason for the whole hardware approach. (3) Conversely, if someone uses the same recipe on a game console or TV and gets over a thousand stars in a week, the path works and you should move fast.

---

Data and sources: GitHub repo page and README (11-06); daily star/fork counts (logged day by day from 10-17; from 10-11 to 10-16 only single-day peaks are available, not daily figures); the author's and reposters' X posts (views and likes as of 11-06); Hackaday's 2026-10-14 coverage; competitor star counts from GitHub (11-06). The traffic brought by Reddit, Hackster, and other platforms is unverified. Whether the App Store approves it and the actual TestFlight install count are unknown.