Providers

AvatarLayer uses a pluggable provider model. Every AI capability — language model, text-to-speech, speech-to-text — is defined by a simple interface. The SDK ships with adapters for popular services, but you can implement any interface to add your own.

LLM providers

All LLM adapters implement the LLMProvider interface:

interface LLMProvider {
  readonly id: string;
  chat(messages: ChatMessage[], opts?: LLMOptions): AsyncIterable<LLMChunk>;
}

The chat method returns an async iterable of LLMChunk objects. Each chunk contains a text delta and a done flag.

Built-in adapters

See LLM Adapters for constructor options and detailed usage.

TTS providers

All TTS adapters implement the TTSProvider interface:

interface TTSProvider {
  readonly id: string;
  synthesize(text: string, opts?: TTSOptions): Promise<Blob>;
}

The synthesize method takes text and returns an audio blob (MP3 by default).

Built-in adapters

See TTS Adapters for constructor options and detailed usage.

STT providers

AvatarLayer supports both batch and realtime speech-to-text. Batch adapters implement STTProvider, realtime adapters implement RealtimeSTTProvider.

Batch adapters

Realtime adapters

See STT Adapters for constructor options, token URL patterns, and detailed usage.

LLM options

Options passed to the chat method or set on the session:

TTS options

Options passed to the synthesize method: