xAI Grok Voice ThinkFast 2.0

This is an example jambonz application that connects to xAI’s Grok Voice ThinkFast 2.0 speech-to-speech API and illustrates how to build a Voice-AI application using jambonz and xAI. The flow uses the same realtime-style llm verb wiring as the other speech-to- speech tutorials, but with xAI-specific configuration required by the vendor.

Authentication

You must have an xAI API key with access to the speech-to-speech models. Provide it in the auth property of the llm verb (or as an application environment variable, as the example does).

1
session.llm({
2
  vendor: 'xai',
3
  model: 'grok-voice-think-fast-2.0',
4
  auth: { apiKey: process.env.XAI_API_KEY },
5
  ...
6
})

Configuring the assistant

Configuration is provided in the form of session_update and response_create client messages inside llmOptions. xAI requires both of them for a session to start properly: audio does not begin flowing until the first session.updated event is received.

1
llmOptions: {
2
  session_update: {
3
    type: 'realtime',
4
    instructions: 'You are a friendly, helpful voice assistant.',
5
    turn_detection: { type: 'server_vad' },
6
    reasoning: { effort: 'high' },
7
    pronunciation: {
8
      replace: [{ from: 'jambonz', to: 'jam bonz' }],
9
    },
10
    audio: {
11
      output: { voice: 'eve' },
12
    },
13
  },
14
  response_create: {
15
    output_modalities: ['audio'],
16
    instructions: 'Greet the caller warmly and ask how you can help today.',
17
  },
18
}

xAI-specific options

The example highlights two xAI-specific knobs that are worth using when you want the assistant to feel more deliberate or more natural on the phone:

• reasoning.effort — set this to 'high' for a more deliberate, analytical response, or 'none' for a faster, more direct reply.
• pronunciation.replace — use this inside session_update to guide the model on how to say custom names, brand terms, or other words that are easy for TTS to mispronounce.

Greeting the caller

Because response_create is required, the assistant can speak immediately after the session is configured. That makes it a good fit for greeting-first flows, IVR handoffs, and other phone experiences where the assistant should begin talking right away.

Interrupting the assistant

xAI handles barge-in server-side. When the caller speaks over the assistant, jambonz flushes queued audio as soon as the interruption event arrives, so the caller can jump in without waiting for the previous utterance to finish.

Events

xAI sends realtime-style server events such as session.updated, input_audio_buffer.speech_started, input_audio_buffer.speech_stopped, and the usual response lifecycle events. Your application can request the ones it cares about in the events property of the llm verb with wildcard support for broader subscriptions.

actionHook properties

Like many jambonz verbs, the llm verb sends an actionHook with a final status when the session completes. The payload includes a completion_reason property indicating why the session ended, such as:

• normal conversation end
• connection failure
• disconnect from remote end
• server failure
• server error