Server & AI
Self-hosting does not keep all voice data local. Cloud ASR, LLM or TTS services still receive the data needed for their work; the data flow depends on your deployment and configuration.
Speech recognition, LLM conversation and speech synthesis all run on the server. The official service is enough to get started; deploy your own server when you want control over voice data, custom models or extensions.
Official service and self-hosting
Which server a device talks to by default is decided by the firmware’s OTA URL. The self-hosted server coordinates voice and model requests; processing can still use cloud providers, depending on each module configuration.
OTA URL, community server projects, deployment and security notes
The capability boundary between device and server
Model and voice configuration
Each community server has its own model and ASR / TTS / VAD configuration entry points; refer to each project’s documentation. A same-field comparison of the four mainstream servers was checked and updated on 2026-11-08.
MCP and extensions
Device-side tool calls and server-side integration
JSON-RPC 2.0 message structure and interaction flow
Endpoint registry and two Home Assistant routes: roles, transports, permissions, status
Go further
- Protocols — WebSocket and MQTT + UDP: how the device talks to the server
- Ecosystem — community servers, clients and ports
- Troubleshooting — service connection and model problems