Start a production voice agent in Voice Agent Builder

Voice Agent Builder is the no-code console for phone and browser agents on Grok Voice. You write how the call should go, attach knowledge and tools, pick a voice, and test it. The July 1, 2026 launch post still lists it as beta.

Build the first agent

1. Sign in at console.x.ai and open Voice Agent Builder (also linked from x.ai/voice).
2. Create an agent. Write a plain-language description of the call: who is calling, what “done” looks like, when to transfer, when to hang up.
3. Upload the documents the agent should retrieve. The launch post lists plain text, Markdown, Word, PowerPoint, Excel, HTML, and JSON. Put them in a collection so more than one agent can share the same policy pack.
4. Attach the tools the call needs: Calendar or Outlook for booking, an API request for order status, web search or X search for public facts, Linear or Notion for tickets, Drive or OneDrive for files.
5. Set guardrails for what it must not say or do (card numbers, off-script topics).
6. Pick a built-in voice, or clone one from about two minutes of audio.
7. Test the change in the browser before you give anyone a number.

Each account includes a free phone number. SIP brings over a number you already have from a telephony provider.

What a call costs

The launch post bills agents at the API audio rate, then $0.05 / min, voices included, no separate platform fee. Telephony on the free provisioned number is an extra $0.01 / min. Confirm the live numbers on the Console pricing page before you size a queue; voice API rates have moved since July.

Every call is recorded and transcribed. You can play the audio, read the transcript, and see which tools ran.

Pitfalls

• Builder is a Console product. A SuperGrok chat plan does not by itself put a number on the public internet.
• Documents in a collection beat pasting the same policy into every prompt. Update the collection when the policy changes.
• Transfer and end-call only work if you attached those tools and wrote when to use them.
• A custom client streams over wss://api.x.ai/v1/realtime. That path is in the Speech to Speech docs.