vercel/ai @ai-sdk/gateway@4.0.0-canary.102

on GitHub

Patch Changes

• a3bb04a: feat(gateway): add experimental realtime model support

  Adds gateway.experimental_realtime() for bidirectional audio/text realtime
  sessions routed through the AI Gateway. Like every other Gateway modality, the
  client speaks the normalized AI SDK realtime protocol and the Gateway
  translates to/from the upstream provider server-side, so GatewayRealtimeModel
  is a thin identity codec. Gateway realtime is server-side only for v0 and throws
  if used in a browser because it returns the resolved Gateway auth token rather
  than a minted ephemeral client secret. Because the browser WebSocket API
  cannot set request headers, the Gateway auth token is carried via the
  Sec-WebSocket-Protocol subprotocol (the same workaround used for OpenAI) and
  the model id rides the ?ai-model-id= query — the WS transport of the
  ai-model-id header used by the HTTP routes. The model id is passed through
  verbatim; the Gateway owns resolution. Provider options (including BYOK) flow
  through the normalized session.update, exactly as they ride the request body
  on the non-realtime routes.

  The versioned subprotocol auth contract is centralized so the client and the
  Gateway server share one definition: getGatewayRealtimeProtocols (client
  encode) and getGatewayRealtimeAuthToken (server decode), plus the
  GATEWAY_REALTIME_SUBPROTOCOL / GATEWAY_AUTH_SUBPROTOCOL_PREFIX constants.

  GatewayProviderOptions documents the stable client-facing option fields while
  remaining open to service-owned options. Runtime validation lives in the Gateway
  service so the server can evolve without requiring an SDK release for every new
  option.