A board that costs about $7 lets an AI agent use a real iPhone. Small app studios could hire it to test every release.
A board that costs about $7 now lets an AI agent see and tap a real iPhone, with no jailbreak and no computer once it's set up. Henry Yuan's open-source ESP32 iPhone Agent turns a cheap real-device tester into a weekend project, and small app studios ship releases without one.
Published Oct 21, 2026
The Wow
What someone built
The ESP32 iPhone Agent mirrors an iPhone's screen to a small ESP32-S3 board through Screen Mirroring and uses the same board as a USB mouse and keyboard with AssistiveTouch on. A companion app sends screenshots to any OpenAI-compatible vision model; a planner and an actor carry out the steps, and every run is kept with its taps and swipes. The README shows it setting a repeating alarm in the Clock app. The repo had about 3,460 GitHub stars on Oct 21, 2026.
- Made by
- Henry Yuan (onehumanbeing on GitHub) (not us)
- Original repository
- ESP32 iPhone Agent repository
- License
- MIT, except the AirPlay receiver code ported from UxPlay, which is GPL-3.0, so the firmware binary is GPL-3.0 (per the README)
- What you need
- An ESP32-S3 board with native USB and PSRAM; the README was tested on a Seeed Studio XIAO ESP32S3 (about $7)
- An iPhone on iOS 18 or later with cellular data (Lightning models need a Lightning to USB Camera Adapter)
- A Mac with Xcode, once, to flash the board and build the app (a free Apple ID can sign it)
- Any OpenAI-compatible vision model API, cloud or self-hosted
- Install
- Flash the board with the repo's flash script, build the iOS app in Xcode, then follow the four in-app setup steps (pair, mirror, pointer, model), per the README.
- A business you could build on it
- Real-iPhone release checks for small app studios
Why it's remarkable
It gives an AI model eyes and hands on the most locked-down consumer device there is, using nothing but ordinary accessibility features, and leaves a replayable trace of every run.
The Opportunity
The business it reveals
Plain-English release checks on a real phone. Indie iOS developers and small agencies often ship updates without anyone walking through the app on a device, because UI tests take time and device clouds cost money every month.
Who would pay
Indie iOS developers and small agencies shipping client apps through TestFlight. Real-device testing is already a paid category (device clouds and QA freelancers).
The Flip
2 ways you could turn this into a business
Real-iPhone release checks for small app studios
- Who pays
- Indie iOS developers and small agencies shipping client apps
- How it would earn
- A fee per release check, or a monthly plan for regular releases (our hypothesis; price untested)
- Why it might work
- Real-device testing is already a paid category, and plain-English tasks remove the test-writing step that small teams skip.
- What already exists
- Device clouds (BrowserStack, Sauce Labs) offer real iPhones by subscription; XCUITest, Maestro and mobile-mcp automate with code or a Mac. This setup is cheaper and needs no scripts, but is slower and shallower.
Fresh step-by-step help guides for iPhone apps
- Who pays
- Support and docs teams at small software companies with an iOS app
- How it would earn
- A fee per guide, plus a monthly refresh when the app changes (our hypothesis)
- Why it might work
- Every run keeps screenshots and actions that can be exported, which is the raw material of a how-to article.
- What already exists
- Guide-capture tools such as Scribe and Tango focus on desktop and web; many teams still take phone screenshots by hand.
Why we'd try "Real-iPhone release checks for small app studios" first: The buyer already knows the pain, the hardware costs almost nothing, and a free five-task trial on their own build is an easy yes.
Build It
How you could build it
A per-release check: a task list from the release notes, real-iPhone runs, exported replays, and a one-page result.
First step: Order a XIAO ESP32S3, follow the README to set it up on your own iPhone, and run three tasks in an app you use every day. Time each one.
Agents and tools
- ESP32 iPhone Agent
- XIAO ESP32S3 board
- A spare iPhone on iOS 18+
- A vision model endpoint
- Your own AI agent to write tasks and reports
Setup
- Flash the board and build the app once on a Mac.
- Dedicate a spare iPhone with cellular data to testing.
- Write a reusable checklist of core journeys (sign up, log in, main action, settings).
- Do two practice runs on public apps to learn the timing and failure modes.
- Make a one-page report template that links each replay.
What the agent does
- Turns release notes into plain-English tasks
- Runs each task on the phone
- Keeps a replay of every step
- Drafts the failure summary
What you do
- Agrees the tasks with the client
- Watches runs and stops them when needed
- Decides which failures are real bugs
- Sends the report and handles test accounts
Your first customer: An indie developer you know, offered five free tasks on their current TestFlight build.
Test demand first: Offer three indie developers a free five-task check. Success: at least one finds a real bug in the replays and asks for the next release. Failure: none of them do.
Your starter instructions
Paste this into the AI agent you use (ChatGPT, Claude, Gemini, Grok Bot or another) as your first message. Copying it starts work in your own agent; it doesn't create or deploy a business by itself.
Running it
What stays human: Writing the tasks with the client; watching the run and keeping the stop button close; deciding which failures are real bugs; writing the report; handling test accounts.
- Startup cost (estimate)
- About $7 for the board, plus an iPhone and a vision-model plan (According to the README, the tested board costs about $7. You need an iPhone on iOS 18+ and a Mac with Xcode once for setup. Model costs per run are not measured)
- Running cost
- Model calls per task and minutes of supervision per release are unknown until measured; the README says each step takes a few seconds, mostly model time.
- How much the agent runs
- People run it; AI helps with tasks.
Pricing your own version? Use the free startup-cost worksheet.
The evidence
What the source shows: The project's README, limitations and license (MIT/GPL-3.0). We read them; we haven't run a client test.
What's ours: Per-release smoke tests for small studios are our application of the kit.
Worth planning around
- Slow by design: the mirrored screen is small (480×1040, 5 fps or less) and each step takes a few seconds.
- iOS doesn't mirror passcode entry, Apple Pay or DRM video, so those flows can't be checked.
- The companion app uses a private iOS API, fine for a self-built app but not for the App Store.
- Mature tools (device clouds, XCUITest, Maestro) go deeper; this wins on setup effort.
Sources
- ESP32 iPhone Agent repository (onehumanbeing/esp32-iphone-agent): README, limitations and MIT/GPL-3.0 license
- Henry Yuan (onehumanbeing) on GitHub
- Seeed Studio XIAO ESP32S3 getting-started wiki
Sources reopened Oct 21, 2026. Spot a mistake? Tell us and we'll correct it in public.
Keep reading
- Services to sell to people building apps with AI (guide)
- How much of the work do AI agents actually run? (guide)
- He still can't write code. His five-person ad agency now runs on about 270 tools he built with Claude. Every owner-run firm needs its first five
- A marketer who doesn't code built an elk-call coaching app with Claude Code and says it hit a $650K run rate in 110 days. Every skill learned by sound is the same opening
- A free n8n workflow flags fake supplier invoices before anyone pays them — and Germany's e-invoice switch makes it a service bookkeepers can sell
- Marc Lou's agent chased a broken-shoe refund through three support desks and got his 850 AED back. Small stores now need a claims desk that can answer agents
- Every discovery and launch plan