Type Alias CallCredentials

Type Declaration

• {
      agentId: string;
      callGen?: string;
      firstMessage?: string;
      instructions: string;
      language?: string;
      provider: "elevenlabs-convai";
      token: string;
      voice?: string;
      voiceToolToken?: string;
  }

  • agentId: string

  • OptionalcallGen?: string

    Generation handle of the call-active window this mint stamped (0.41.0). Self-plumbed integrations: pass it to endSession({ callGen }) at call end so the server clears exactly this call's window — a stale handle's late teardown then can't clear a newer call's window. connectCall() handles this automatically.

  • OptionalfirstMessage?: string

    First-message override ('' = suppress the canned greeting so the call opens in-character / context-aware). Pass as overrides.agent.firstMessage.

  • instructions: string

    Pass as overrides.agent.prompt.prompt when starting the EL session.

  • Optionallanguage?: string

    Server-resolved agent language (ISO 639-1, e.g. 'zh'). Pass as overrides.agent.language so the call opens in the user's language.

  • provider: "elevenlabs-convai"

  • token: string

    EL conversation token (open WebRTC with @elevenlabs/client).

  • Optionalvoice?: string

    Server-resolved EL voice id (the user's cloned/localized voice). Pass as overrides.tts.voiceId so the call isn't a stock default voice.

  • OptionalvoiceToolToken?: string

    Call-scoped bearer the agent's memory webhook tool presents back to us mid-call (0.45.0). Self-plumbed integrations: pass it to Conversation.startSession under BOTH dynamic variable names (0.47.0) — dynamicVariables: { [VOICE_TOOL_VARIABLE]: creds.voiceToolToken, [VOICE_TOOL_VARIABLE_PLAIN]: creds.voiceToolToken }. ElevenLabs IGNORES a client-supplied secret__ variable, so the plain twin is the one that arrives; it reaches the tool as a request-body property the platform fills, never as prompt content the LLM sees. Passing only the first name raises no error — it just leaves the companion unable to recall anything. connectCall() handles this automatically.

    ABSENT, not empty, when the session cannot read memory: a client with no memory grant has nothing to pass, rather than something that fails inside a live call. Omitting it does not break the call — recall just answers "no stored memory".

• {
      callGen?: string;
      clientSecret: string;
      expiresAt: number | null;
      model: string;
      provider: "openai-realtime";
      voice: string;
  }

  • OptionalcallGen?: string

    See the elevenlabs-convai branch — same window-generation handle.

  • clientSecret: string

    Ephemeral client secret (ek_…); open WebRTC directly to OpenAI.

  • expiresAt: number | null

  • model: string

  • provider: "openai-realtime"

  • voice: string