Grok Voice Agent Builder

What the product is

Grok Voice Agent Builder is a no-code voice agent builder from xAI, aimed at support and sales teams, operators and developers who want agents that answer phone calls and hold live spoken conversations. The directory entry at theaidude.net calls it "xAI's no-code platform for building, testing, and deploying real-time voice agents" and says it launched on July 1, 2026. It is a voice product first: a chatbot builder in the sense that the agent converses, but the conversation happens out loud, over a phone line or a browser session, rather than in a text widget.

Everything below is attributed to the third-party pages that describe the product: reviews, comparison listings and a developer tutorial. Wikipedia lists xAI as a company with more than 1,200 employees in 2025, a figure it says includes over 900 hourly paid AI tutors. Wikipedia also reports that xAI became a subsidiary of SpaceX in February 2026 and was rebranded SpaceXAI in July, and names SpaceX as its parent.

What you build and how

The editor. Slashdot and SourceForge both carry the same product description. It says users can "write a plain-language description of how calls should flow", attach documents, connect tools, set guardrails and get to a working agent, and it puts the setup time at under 2 minutes. The Rundown AI describes the beta builder as a no-code interface for configuring the agent, its knowledge, tools, guardrails, voice, phone access and call review. A review on Cobus Greyling's Substack gives an example of the kind of prompt involved: a friendly support agent that greets callers, asks for a name and inquiry, answers from uploaded docs and books discovery calls. The same reviewer adds a limit worth knowing about. Flow building and definition, he writes, is "only done on a high level", so this is not a node-by-node flow chart editor.

Knowledge. SourceForge and Slashdot say the agent retrieves from uploaded knowledge bases in plain text, Markdown, Word, PowerPoint, Excel, HTML and JSON, among other formats. The Rundown AI lists retrieval from uploaded knowledge as one of the things a deployed agent can do mid-call.

Tools and logic. According to The Rundown AI, an agent can invoke business functions, search the web or X, and transfer a caller to a person. Slashdot's listing names built-in web search, file search and computer tools plus custom function-calling tools, and marks Model Context Protocol (MCP) support as Yes. DataCamp's tutorial shows the custom function loop in practice: the model sends a function call event with arguments, your code runs the function and returns the output, and you then send a response.create event so the agent keeps talking. Skip that last step, DataCamp warns, and the agent goes silent. Its demo registers functions named book_appointment, transfer_to_human and end_call, and it says built-in tools such as web search, X search, collections search and remote MCP tools are available alongside them. The Rundown AI is blunt about the work this implies: teams need verified tool permissions, consent and recording disclosures, escalation paths, adversarial testing, monitoring and cost controls.

Phone and publishing. The Rundown AI says the builder includes a phone number and that developers can connect their own carrier or PBX over SIP. It also lists phone keypress support and human call transfer. A review on medux.io describes provisioned phone numbers, SIP for existing numbers and carriers, native telephony codecs, WebSocket clients and human transfer. The Substack review adds that provisioning happens in the console and that xAI handles the carrier-side details for free numbers. For a custom setup you configure SIP trunks or inbound routing rules. eesel.ai says you can also test an agent in the browser without a phone.

API and model. Slashdot's listing marks API access as Yes, and The Rundown AI says custom integrations can still use the API. AlphaSignal, a news site, reports that xAI built the whole voice stack in-house, training its own voice activity detection, tokenizer and audio models. The platform's pitch, per SourceForge, is one interface on a speech-to-speech path rather than separate speech-to-text, language model and text-to-speech services stitched together.

Languages and speed. The Substack review repeats the vendor's promises, which it summarizes as a voice agent in 2 minutes, sub-second latency and 25+ languages, and treats them as promises rather than measurements. eesel.ai says Think Fast 2.0 scores 82.9% on the Artificial Analysis speech-to-speech index and answers in 0.70s. AlphaSignal notes that benchmark numbers of this kind are self-reported, though it says the tau-voice benchmark is independently run by Sierra.

Operations. The Substack review says the platform scales automatically and can handle dozens of simultaneous calls without separate instances. SourceForge and Slashdot describe observability as part of the package, and The Rundown AI lists call review in the builder.

What the reviews say about pricing and limits

The figures below are what reviewers report, and several of them are dated.

• Per-minute pricing. The AI Dude lists a single usage-based plan "from $0.05/min" with no free tier documented, and says its pricing was last edited on July 2, 2026 and not re-verified since. AlphaSignal reports $0.05 per minute of audio with voices included and no separate platform fee. DataCamp says audio bills at $0.05 per minute counting both what you send and what you receive.
• The newer model. The Rundown AI reports that xAI's pricing documentation lists Grok Voice Think Fast 2.0 at $0.08 per audio minute plus $0.004 for text input, and marks the older Think Fast 1.0 price of $0.05 per minute as deprecated. eesel.ai calls that 60% above the 1.0 rate and says the default alias moved to 2.0 on August 5, 2026. A reader budgeting today should use the $0.08 figure.
• Telephony. AlphaSignal, eesel.ai, medux.io and DataCamp all give $0.01 per minute extra for a provisioned phone number. eesel.ai works one example: a 5-minute support call on a provisioned number is about 30 cents in voice and telephony, before tool calls.
• Tool meters. eesel.ai lists web and X search at $5 per 1,000 calls and document search at $2.50 per 1,000. DataCamp says text input events bill at $0.004 each, and that function_call_output results and response.create events are not billed. The Rundown AI adds that telephony, search tools, storage and other services can add charges.
• Free access. eesel.ai says there is no published free tier on the pricing page, but each account includes a free phone number for a first test call.
• Limits. Medux.io and DataCamp both report up to 100 concurrent sessions per team and a 120-minute maximum session, which medux.io says is subject to account limits. DataCamp adds that resumed session history is dropped after 30 minutes of inactivity.
• Compliance. DataCamp tells teams handling patient data to read xAI's compliance terms carefully. None of the reviews names a certification.
• Status. The Rundown AI calls the Voice Agent Builder a beta, and eesel.ai describes sentiment among builders as cautiously positive for a beta from a lab that ships fast.

Who it suits and who it does not

The AI Dude says the builder is aimed at support and sales teams deploying phone-based agents for support calls, appointment booking and outbound lead qualification. DataCamp frames the same territory from the developer side: use voice when the interaction happens live and the agent needs to act, such as appointment booking, customer support and internal lookup workflows. A small team that would otherwise stitch together separate speech, language and voice vendors, and wants one per-minute meter, is the clearest match.

It suits less well when a text chatbot would do. DataCamp lists that case, batch transcription, workflows not yet tested with real users, and teams that cannot yet handle errors, privacy and escalation safely as situations to avoid. If you want a conventional text chatbot for a website, a dedicated chatbot builder is the closer tool.

There are real caveats. The Substack author expects organizations to outgrow the builder as their applications grow, and notes that some teams will want their own harness and independence from any one model. The high-level flow definition means complicated branching logic lives in the prompt and in your own functions, not on a canvas. The 100-session ceiling matters to a busy contact center. And because the pricing reports above are dated, a buyer should confirm current rates, since eesel.ai's reporting shows the headline per-minute price changed within weeks of launch.