DIY voice AI hardware that outgrew its repository

What it does

Starmoon is an open-source hardware and software kit for building a handheld voice AI companion. An ESP32-S3 board, INMP441 microphone, and MAX98357A amplifier live inside a 3D-printed enclosure only slightly larger than an Apple Watch, streaming speech to GPT-4o, Deepgram STT, and Azure TTS while a Next.js frontend handles character management and a trends dashboard. The firmware, software stack, and case files are all GPLv3-licensed, and the project claims local self-hosting is possible if you want to keep conversations off third-party clouds.

The interesting bit

The README’s first word is effectively “deprecated”: the maintainers have frozen Starmoon and continued development as ElatoAI, citing improved WiFi, better two-way audio reliability, and a production-ready architecture. That makes this repository less of a product and more of a fully documented snapshot of how to bolt a cloud LLM onto an ESP32 with a soldering iron.

Key highlights

• Complete BOM-to-firmware kit: off-the-shelf ESP32-S3, INMP441 microphone, MAX98357A amplifier, and a 3D-printed case.
• Swappable “AI characters” and a web-based trends dashboard for conversation analytics.
• Designed to reduce screen time by replacing app-based chatbots with a dedicated physical gadget.
• Explicitly deprecated; active development has migrated to the ElatoAI fork.

Caveats

• The project is officially deprecated; active development continues under ElatoAI, so don’t expect fixes in this repo.
• The default setup asks for a Starmoon cloud API key, and while the README mentions local self-hosting, it never spells out how to disconnect from its own cloud entirely.
• Hardware audio streaming depends on a 2.4 GHz WiFi connection; 5 GHz is not supported.

Verdict

Tinkerers and hardware hackers who want a reference design for pocket voice AI will find a complete parts list and working firmware. If you actually want to ship a product or need reliable support, the README itself points you to ElatoAI instead.