esp32agent

esp32agent talks to an ESP32 Agent board plugged into the iPhone. The board is the phone's USB mouse and keyboard and receives its AirPlay screen mirroring; this computer reaches it over Bluetooth. Every command prints JSON: {"status":"ok","data":...} or {"status":"error","error":"..."}. The MCP server (esp32agent mcp) offers the same as tools: screenshot, tap, longpress, swipe, type_text, press_key, press_button, device_info, plus history, intent_list, intent_show and intent_save.

Setup (once)

• esp32agent pair within 3 minutes of plugging the board into the iPhone.
• Screen Mirroring on the iPhone: Control Center > Screen Mirroring > ESP32 Agent.
• esp32agent calibrate (also runs by itself on the first tap). It needs the mouse pointer visible on screen.

The loop

1. esp32agent screenshot (writes a PNG in the current directory and prints its filePath; read the image). It waits until the screen has settled after your last action, so it never shows the screen from before it.
2. Decide one action from what you see.
3. Act, with coordinates in pixels of that screenshot (full size: don't pass --scale/--max-size if you take coordinates from it):
   • esp32agent io tap X,Y (add --aim after a missed tap: finds the pointer and corrects first; --count 2 double-taps, e.g. to zoom a map or select a word)
   • esp32agent io longpress X,Y --duration 1500
   • esp32agent io swipe X1,Y1,X2,Y2 (scroll a list: swipe up from the lower to the upper middle)
   • esp32agent io text 'hello', esp32agent io keys enter, esp32agent io keys cmd+a
   • esp32agent io button HOME (Home Screen)
4. Screenshot again and check the result before the next step. Never chain several taps blind.

Typing

• The board types printable ASCII only, one key per character.
• Chinese is typed as pinyin: io text '拿铁' types natie (the reply says typed_as). With the iPhone's Chinese (pinyin) keyboard active, commit the first candidate with io keys space, or screenshot and tap the right candidate. Many search boxes accept the plain pinyin as well.
• The board is a hardware keyboard, so iOS may hide the on-screen keyboard; the text field still takes the keys.
• Tap the text field first so it has focus; io keys cmd+a then io keys backspace clears it.

Save what worked

After a task succeeded, save it so the next run is faster:

1. esp32agent history show latest (MCP: history with action show): the run's steps in 0-1000 screen units, numbered, each with the screenshot it was done on. Every action and screenshot is recorded by itself; a new run starts after 10 minutes idle, or with esp32agent history new before a new task.
2. Pick the steps that did the job (skip detours, misses and retries).
3. esp32agent intent save --name 'Alarm 7:00' --prompt 'Set an alarm for 7:00 AM every day' --hint 'Type the time instead of scrolling the wheel' --from-run RUN --steps 2,3,5 (MCP: intent_save with name, prompt, hints, from_run, steps). One short, clear instruction as the prompt and at most 6 concrete hints. Only replayable moves can be steps: tap, swipe, type, key, home. The same name replaces the old intent.

Run a saved intent

1. esp32agent intent list, then esp32agent intent show NAME (MCP: intent_list, intent_show): the prompt, hints and recorded steps, each with its reference screen (a JPEG path).
2. Follow the steps as references, not blindly: coordinates are 0-1000 units of the screen, so convert them to pixels of your screenshot (x * width / 1000, y * height / 1000).
3. Take a screenshot after each step and check it looks like the next step's reference screen.
4. On a mismatch (another screen, a popup, a changed layout) drop the recorded steps and go on with the normal loop, using the prompt and hints.

Errors to act on

• busy: the app is using the phone: the iOS app is driving right now. Wait and retry later; don't fight it.
• screen mirroring is off: ask the user to start Screen Mirroring to ESP32 Agent.
• video frozen: ask the user to stop and restart Screen Mirroring.
• not paired: ask the user to replug the board and run esp32agent pair within 3 minutes.
• phone usb asleep: the phone is locked; ask the user to unlock it.
• not supported on ESP32 Agent: there is no app list, UI tree, file access or webview; work from screenshots.

Limits

• Video comes over Bluetooth: a screenshot can be up to a second behind, animations look choppy.
• Only HOME as a button; no volume, power, or app install/launch (open apps from the Home Screen or Spotlight: io button HOME, swipe down in the middle, io text 'Settings', io keys enter).