Which package?

Map

UI / App
   ↕ events
@ottervoice/core          ← session, state machine, VAD, barge-in, usage
   ↕ RuntimeAdapter
runtime-web | react-native | node
   ↕ Provider interfaces
provider-deepgram | elevenlabs | openrouter | azure-speech | custom

Choose

Package	When
@ottervoice/core	Always. createVoiceSession / createOtterVoiceSession, events, mocks.
@ottervoice/runtime-web	Browser mic + playback.
@ottervoice/runtime-react-native	Expo / RN with injected native audio.
@ottervoice/runtime-node	CLI / server audio and networking.
Providers	Pick a cloud vendor — or implement the same interfaces.
@ottervoice/protocol	Forward session events across processes.
@ottervoice/provider-utils	Custom providers: credentials, SSE, WebSocket ASR.

One client contract, two backend implementations

Backend	Inside AudioLLMProvider	Client difference
Native	audio in → native speech model → text + audio out	None
Composite	audio in → ASR → LLM → TTS → text + audio out	None

VoiceSession always receives providers.audioLlm, plus optional providers.asr for separate captions. Model count and vendor selection stay behind the provider or trusted gateway.

Bundled OpenRouter adapters are optional demos, not a fixed production route.