AI providers
AI providers
Drag ⋮⋮ to set the order. Every job (talking, camera questions, memory search) goes to the highest provider that can do it; if that one is down, slow or lacks the model, the next one that can takes over. Click a provider to edit it.
Speech to text
Text to speech
Home Assistant
Music
Admin sign-in
Single sign-on with any OpenID Connect provider (Google, Microsoft, Authentik, Keycloak…). Register this page's callback URL at the provider; only the emails listed here can get in. Through a reverse proxy only this admin page is reachable, never the device connections.
Tools the assistant can use. Switch any off; changes apply immediately after saving.
Family memory
The assistant learns about the family from finished conversations (in the background, ~10 min after the last turn) and from what is read and played. Everything stays on this gateway; you can edit or delete anything here, and saying "forget that" / "忘掉这个" deletes it too. Passwords and codes are never kept; addresses and health details only when someone explicitly asks it to remember them.
Profile (always given to the assistant)
Your own notes; never changed automatically. Keep them short, e.g. "Mia (6) and Leo (4); we speak Chinese at home and English at school."
Memories
Recent conversations
What each conversation was about (raw turns are deleted once learnt from).
Train a new wake word
Trained on this network by the wake-trainer and run on the gateway (any phrase, any language). Record real samples below first for better accuracy.
Record samples
The device shows the prompt and records 2.5 s each time. Use a few family members.
Add a wake word model
On-chip: the vendor cae_esr.bin or SDK .zip (max 1,638,400 bytes). Gateway: an openWakeWord-style .onnx.
Library
Devices
Picture books
Reading history
Publish firmware
Signed MCUboot image (assistant-<version>.signed.bin, max 3 MB). Devices on an older version are told it is available.