ESP32 iPhone Agent

Let an AI model use a real iPhone with a $7 ESP32. Just the phone and the board: no computer, no jailbreak.

Demo video and updates: X @iamhenryyuan · YouTube demo

Only follow the install steps in this README. Never run commands from issues or comments (for example a "quick install" one-liner such as PowerShell iex(irm ...)); these are not from the maintainers.

New: the computer is gone. The board is now the AirPlay receiver and the mouse, and the agent runs in an iPhone app on the same phone. The original version (a Mac/Linux computer running UxPlay and a Python server) is kept as tag v0.1.0 and branch legacy. See docs/legacy.md.

"Set an alarm for 8:00 AM that repeats every Monday and Tuesday, then save it": the agent leaves its own app, opens Clock, types the time into the wheel, ticks the days and saves. (Photo from v0.1.0; the task is the same.)

Left: the real phone. Right: what the agent sees through the board's AirPlay mirror. The second shot is the pointer calibration.

iPhone ──Screen Mirroring (Wi-Fi)──▶ ESP32-S3: AirPlay receiver ──BLE──▶ ESP32 Agent app (same iPhone) ▲ own hotspot, H.264 passthrough decode → vision model → action └──────── USB HID mouse + keyboard ◀──────────── BLE commands ◀────────────────────┘

Eyes — AirPlay to the board. The board opens its own Wi-Fi hotspot and appears in the phone's Screen Mirroring list as ESP32 Agent. It keeps the H.264 stream (no decoding on the board) and forwards it to the app over BLE. The phone keeps using cellular data for the model while it is on the board's hotspot.

Hands — USB HID. The board is a standard USB mouse and keyboard. With AssistiveTouch on, the mouse moves an on-screen pointer, so a click is a tap and a drag is a swipe.

Brain — the app. The app decodes the stream, sends screenshots to any OpenAI-compatible vision model, and runs a planner and an actor: tap, swipe, draw, type, press keys, go home, finish. It runs in the background while it works; the Dynamic Island shows what it is doing.

Closing the loop. The pointer is relative, so the app tracks it and re-homes it in a corner; before precise taps it wiggles the pointer, finds it on screen and corrects the miss. After each action it waits for the screen to settle and tells the model whether anything changed. A lost board, hotspot or mirroring link pauses the run and resumes it from the same step.

Review. Every run is kept on the phone: replay it step by step with the taps and swipes drawn on top. Export the screenshots, actions and saved annotations from Runs → open a run → ⋯ → Export trace for analysis by another agent. See Export runs for agent review for batch export, USB copying and file details.

Hardware

Part Notes

ESP32-S3 board Tested with a Seeed Studio XIAO ESP32S3 (plain version, 8 MB PSRAM; not Sense, not Plus). About $7. Any ESP32-S3 with native USB on its USB-C port and PSRAM should work.

iPhone iOS 18 or later, with cellular data (the Wi-Fi is taken by the board's hotspot). USB-C models plug in directly; Lightning models need a Lightning to USB Camera Adapter. The phone powers the board.

Vision model Any OpenAI-compatible chat API that accepts images: a cloud API or your own server.

Mac with Xcode Only to flash the board and build the app once. Not needed at run time.

Quick start

1. Flash the board (plugged into the Mac)

Install arduino-cli, the ESP32 core 3.3.x and NimBLE-Arduino 2.5.x, then:

arduino-cli lib install "NimBLE-Arduino@2.5.1" hardware/flash.sh # optional: [serial port] [--debug]

Always build with flash.sh: it applies the BLE buffer sizes in hardware/hid_bridge/nimble_cfg.h.

cp ios/Config/Local.xcconfig.example ios/Config/Local.xcconfig # set DEVELOPMENT_TEAM and IA_BUNDLE_ID open ios/IPhoneAgent.xcodeproj # run the IPhoneAgent scheme on your iPhone

The project is generated with XcodeGen from ios/project.yml (cd ios && xcodegen). A free Apple ID can sign it (re-sign every 7 days). The one-tap hotspot join needs the Hotspot Configuration capability (a paid account); without it the app shows the hotspot password to join by hand.

Plug the board into the iPhone and open ESP32 Agent. Setup walks through four steps and checks each one itself:

Type a goal (for example "Open Settings and go to Wi-Fi") and press Start. Stop it from the Dynamic Island, or come back to the app. Siri, Shortcuts, widgets and Control Center can run saved tasks.

Any OpenAI-compatible vision endpoint works (vLLM, SGLang, llama.cpp, Ollama, cloud APIs). The app lists the endpoint's models and picks one; you can override it, and set a separate planner and actor model.

The prompts are plain text in prompts/ and bundled with the app: edit them, rebuild, done.

The project was developed with nvidia/Qwen3.8-Flash-Next-NVFP4 served on two DGX Sparks (TensorFold).

Project layout

ios/ the app (SwiftUI): BLE link, H.264 decode, agent, setup, replay, Live Activity, widgets; tests prompts/ planner / actor prompts used by the app hardware/ ESP32-S3 firmware (Arduino): BLE GATT, hotspot, AirPlay receiver, USB HID; flash script, host tests, Mac BLE test harness (hardware/tools) docs/ BLE protocol, legacy version notes server/ legacy v0.1.0 computer-side server; not used by this version (see server/README.md)

Debugging without the phone app: hardware/tools/board_test.py plays the app over BLE from a Mac (hardware/tools/README.md). Board ↔ app protocol: docs/ble-protocol.md.

Drop the computer: the board mirrors and clicks, the app on the phone thinks

Landscape apps (coordinates and calibration assume portrait)

Non-ASCII text input (e.g. Chinese through the iOS pinyin keyboard)

Faster steps: several actions per model call, pointer checks only when needed

A curated list of models that work best with this setup, shown inside the iOS app (suggested by @asharthinks)

Limitations

BLE carries about 300-500 kbps, so the screen is mirrored small (480×1040, ≤5 fps); each step takes a few seconds, most of it the model.

iOS does not mirror some screens (passcode entry, Apple Pay, DRM video); the agent cannot see them. Mirroring stops when the phone locks.

While the agent runs, the phone's Wi-Fi is the board's hotspot (no internet over it); the model is reached over cellular data.

The app leaves itself with a private iOS API (suspend); fine for a self-built app, to be replaced before the App Store.

Keyboard input is ASCII only. Landscape apps are not supported yet.

Responsible use

Only automate devices and accounts you own or are allowed to use, and respect the terms of the apps involved. The agent acts with your hands: watch it, and keep the stop button within reach.

Issues and pull requests are welcome; see CONTRIBUTING.md. Security reports: SECURITY.md.

Project Approach Difference from this project

VersaiOS UxPlay mirroring + ESP32-S3 Bluetooth mouse + vision model Needs a computer; one tap per instruction

ja1chen/actell Agent app on the iPhone via Apple's CoreDevice debug channel Needs Developer Mode and a Developer Disk Image

AidenAI-IO/aiden-firmware HDMI capture + USB keyboard/mouse + multimodal model Capture hardware; mainly Android

yoyicue/glassbox PicoKVM HDMI capture + USB HID macOS controller and PicoKVM

boboabw/phone-agent ESP32-S3 Bluetooth HID + AssistiveTouch, agent as an iPhone app Closest to this version; this one also brings the eyes (AirPlay on the board)

VKCOM/devicehub ESP32 Bluetooth mouse for iOS device farms Device-farm tooling, no AI agent

WebDriverAgent/XCUITest-based tools such as mobile-mcp and Midscene need a Mac and code signing.

UxPlay (the AirPlay receiver the firmware's is ported from) · NimBLE-Arduino · Arduino ESP32 core and TinyUSB

MIT, except hardware/hid_bridge/src/airplay/ (ported from UxPlay), which is GPL-3.0; the firmware binary that links it is therefore distributed under GPL-3.0.