Reprogram a $50 ESP32 AI pendant into a talking pocket gadget

Hollister bought a Waveshare ESP32-S3 1.75C, a small round touchscreen pendant that Meta lists as a starting point for gadgets that talk to its Muse AI cloud. He plugged it into his PC and pointed Claude at Meta's sample code on GitHub to flash and pair it. The stock build was slow, silent and showed only a few words of each reply, so over a few Claude Code sessions he plugged the board back in and described each feature he wanted. Claude Code added a swipeable conversation history, spoken replies through an ElevenLabs voice, a remote shutter with live preview for his phone camera, a Bluetooth preview, a set of avatars, a care-for-your-pet minigame with levels of hunger, boredom and dirt, a three-hand watch face, and a video player loaded with trimmed clips it found and downloaded.

He says almost every gap in the stock build got filled. The watch face took about 15 minutes. A bug where releasing the talk button too early sent near-empty audio was fixed within minutes by only sending clips longer than one second. The remaining lag comes from Meta's servers. He never got it to discover smart home devices on its own, and he says he cannot judge whether the code is any good.

The author has no coding experience and kept the device on guest Wi-Fi because he does not trust the Muse service with his network. The ElevenLabs voice came from its free tier. This is press reporting of the writer's own project, published by The Verge.