How a $50 Open Devkit and Claude Code Recreated Meta’s Upcoming AI Charm

Meta is preparing to enter the standalone consumer hardware space this holiday season with the Muse Charm, a pocket-sized AI helper designed to act like a modern, interactive Tamagotchi. However, tech journalist Sean Hollister demonstrated that users do not necessarily need to wait for Meta's retail device. Using a $50 development kit and Anthropic's Claude Code, he assembled and expanded a working DIY companion pendant at home.

The $50 Hardware Base

The project relies on the Waveshare ESP32-S3 1.75C, a $50 pocket watch-style device featuring a 1.75-inch touchscreen. The push-to-talk pendant devkit is among the hardware options officially recommended by Meta for building "Muse Gadgets"—devices running Meta’s open-source code and authentication tokens to connect directly to the company's Muse AI cloud.

While the hardware physically resembles Meta’s planned pendant, its out-of-the-box software proved limited. Initially, Meta’s setup interpreted push-to-talk speech as transcription rather than prompt commands. The default configuration did not produce spoken audio responses, displayed only a few words at a time on its tiny screen without conversation logs, and featured only a single static "Jolly" avatar. Furthermore, unlike consumer smart speakers from Amazon or Google, it lacked automatic local smart home discovery, and sample camera streaming code from Meta software engineers Anant Narayanan and Jeremy Payne lacked a remote shutter function.

Rapid Feature Expansion via "Vibe Coding"

Despite having no formal programming background, Hollister used Claude Code to overhaul the pendant's firmware by connecting the device to a PC and prompting the model to write and modify code.

Through iterative prompts, Claude resolved several usability and feature shortcomings:

• Audio and Display Improvements: Claude created a swipable conversation history screen to review full text logs and integrated a free-tier British text-to-speech voice from ElevenLabs so the device could answer aloud. It also patched a push-to-talk bug by filtering out audio clips under one second in length.
• Remote Camera Controls: The model developed an interactive remote viewfinder for a paired smartphone over both Wi-Fi and Bluetooth, corrected an inverted front-camera feed, and added a functioning shutter button.
• Virtual Pet Mechanics: Hollister had Claude generate multiple avatars inspired by Meta Connect demos and built an interactive Tamagotchi-style minigame. The software tracks four tiers of hunger, boredom, and cleanliness, penalizing overfeeding with messes and excessive play with fatigue.
• Utility Additions: In roughly 15 minutes, Claude generated a working analog pocket watch face with three animated hands, alongside a dedicated video player configured to play clipped video files on the 1.75-inch screen.

Devkits vs. Dedicated AI Gadgets

The experiment showcases the accessibility of "vibe coding"—using conversational AI agents to rapidly produce working embedded software on low-cost hardware. At the same time, Hollister highlighted limitations in Meta’s current cloud pipeline, noting inconsistent response latencies from Meta's servers.

The project raises broader questions about the appeal of specialized AI hardware. Without extra features and customization, the pendant offered little utility beyond a standard smartphone running the Muse mobile app. For tinkerers, open development kits combined with code-generating LLMs offer a flexible alternative, leaving an open question as to whether locked-down commercial AI pendants will provide enough unique value to justify their existence.