xiaozhi-esp32 brings AI voice interaction to ESP32 hardware devices using large language models. It uses a dual MCP architecture for device-side peripheral control and cloud-side capability extension.
Project overview
It combines offline voice wake-up on ESP32 with LLM-driven ASR and TTS pipelines, and extends control to peripherals and cloud services through MCP.
Project type
MCP
Deployment
Refer to project documentation
License
MIT
Best for
Developers and creators building AI agent automation on supported ESP32 chip platforms (ESP32, ESP32-C3, ESP32-C5, ESP32-C6, ESP32-S3, and ESP32-P4).
Key capabilities
Provides offline voice wake-up, ASR + LLM + TTS pipelines, and realtime end-to-end voice models.
Enables control of device peripherals (Speaker, LED, Servo, GPIO) via the MCP protocol.
Extends large model capabilities to smart home control, PC desktop operation, knowledge search, and email via cloud MCP.
Supports Wi-Fi, wired Ethernet, USB RNDIS, and 4G networking (ML307/EC801E or NT26 Cat.1).
Identifies the current speaker using 3D Speaker.
Supports OLED/LCD displays with emoji and camera vision input on supported boards.
Manages battery display and device power.
Allows customization of wake words, fonts, emojis, and chat backgrounds via a web-based editor.
Limitations and risks
ESP-IDF v5.5 is retained only for legacy hardware paths.
Inference requires external large models like Qwen or DeepSeek. The optional xiaozhi.me server can also be used.
Getting started
Beginners can flash pre-compiled firmware without setting up a dev environment. The documented first success path involves choosing a supported ESP32 board, following the beginner's firmware flashing guide, flashing firmware to the device, connecting to Wi-Fi, and registering a personal account on xiaozhi.me to use for free.