July 2026
sihah-esp32
A voice-controlled AI chatbot on ESP32 — offline wake word, streaming speech to an LLM and back, and device control over MCP. My build of the open-source XiaoZhi project.

Not my design — this is me getting hands-on with on-device voice AI and MCP by building and customising an existing open-source project, then learning how every piece fits.
A pocket-sized voice assistant on an ESP32. You wake it with a hotword, talk to it, and it answers — speech in, LLM in the middle, speech out — and it can control the device and cloud services over the MCP protocol.
Credit where it's due: this is my build and customisation of the open-source XiaoZhi ESP32 project (
xiaozhi.me), not an original design. I flashed it, wired the hardware, and dug through the code to understand how on-device voice AI actually works.
What it does
- Offline wake word via ESP-SR, then streaming ASR → LLM → TTS (Qwen / DeepSeek).
- MCP on the device (control speaker, LED, servo, GPIO) and MCP in the cloud (smart-home, search, email).
- OLED / LCD display with emoji, battery management, and multi-language support (English, 中文, 日本語).
- Runs on ESP32-C3 / S3 / P4 boards.

What I got out of it
Voice AI on a microcontroller is a stack I'd never touched: audio codecs (OPUS), I2S mics, wake-word models, and a real-time ASR/LLM/TTS pipeline squeezed onto a chip. Building someone else's project end-to-end — and getting it running on my own hardware — was the fastest way to actually understand it.
More projects
view all →Quack prototype
A pixel-duck agent UI for the ESP32 Cheap Yellow Display — idle / listening / thinking / talking sprites, an I2S mic, and touch pages — prototyped in the Wokwi simulator. The precursor to ohmyclawd.

ohmyclawd
A Claude Code usage monitor on a tiny ESP32 desk display — session and weekly usage as animated pixel art, with a retro clock. My macOS + cloud fork of an open-source project.

Canopy
An agent situation-room: a panel of AI leads debates a call across your projects, then you approve it. Bounded, mediated, human-approved.