explainx / blog
xAI launched Voice Agent Builder in beta β a no-code platform for production Grok Voice agents with telephony, MCP, guardrails, and observability at $0.05/min. Benchmarks, pricing, and how it compares to stitched STT+LLM+TTS stacks.
May 4, 2026
AI agents that place real calls and texts need carrier identity, routing, and auditβnot a consumer SIM. Saperly packages lines, modes, and docs around agent workflows; here is what the product actually lists.
Jul 3, 2026
Starting July 6, Tesla limits each employee to $200 per week on AI coding tools β with manager sign-off for more. It is the latest enterprise pullback after unchecked usage spiked bills. Is $200/week generous or a verdict on LLM ROI?
Jun 30, 2026
@XDevelopers announced hosted Model Context Protocol servers so Grok Build, Cursor, Claude, and VS Code can call the X API with your account permissions. Two endpoints, one local xurl bridge, zero custom server setup.
On July 1, 2026, xAI announced Voice Agent Builder in beta: a no-code platform to configure production voice agents on Grok Voice β telephony, knowledge bases, tools, MCP connectors, guardrails, and call observability in one place. Pricing starts at $0.05/min of audio with voices included; a free phone number ships with every account.
The pitch is familiar to anyone who has shipped phone agents: most stacks stitch three APIs β speech-to-text, a language model, and text-to-speech β often from different vendors. Each hop adds latency, cost, and failure modes. xAI's answer is a single speech-to-speech path built for Grok Voice, exposed through an operator console rather than a bag of SDKs.
| Question | Answer |
|---|---|
| When did it launch? | July 1, 2026 β beta at console.x.ai/voice/agents |
| What does it cost? | $0.05/min audio (API rate, voices included) + $0.01/min on free provisioned telephony |
| Do I get a phone number? | Yes β free number per account; SIP import or WebSocket client also supported |
| Is it really no-code? | Plain-language agent prompt + document uploads + tool/MCP wiring; xAI claims ~2 minutes to first agent |
| MCP support? | Yes β alongside API tools, calendars, X search, Linear, Notion, Drive, OneDrive |
| Benchmark claim? | Ο-voice Bench: Grok Voice Think Fast 67.3% vs Gemini 3.1 Flash Live 43.8%, GPT Realtime 1.5 35.3% |
| vs GPT-Realtime-2? | OpenAI = model API; xAI = platform + telephony + ops on Grok Voice β see comparison below |
Voice agents that handle real phone calls face conditions demo videos skip:
xAI says Grok Voice was trained on that traffic. They publish Ο-voice Bench scores under those conditions:
| Model | Overall (Ο-voice Bench) |
|---|---|
| Grok Voice Think Fast 1.0 | 67.3% |
| Gemini 3.1 Flash Live | 43.8% |
| GPT Realtime 1.5 | 35.3% |
Category breakdowns (retail, airline, telecom) appear on x.ai/voice. Treat vendor benchmarks as directional β run your hardest call flows before production β but the gap xAI quotes is large enough to force a bake-off if you are re-platforming support lines.
xAI's setup flow:
Actions during calls: lookup records, change CRM rows, transfer to human, end call, and real-time notifications so a team can see tool use and intervene.
After calls: recordings, transcripts, and a trace of which tools fired. Guardrails block off-script topics or sensitive reads (e.g. reading back full card numbers).
| Line item | Rate (xAI, July 2026) |
|---|---|
| Grok Voice audio (agent runtime) | $0.05 / min |
| Voices | Included (no separate TTS meter) |
| Platform fee | None (per xAI announcement) |
| Telephony (free provisioned number) | +$0.01 / min |
xAI contrasts this with stacks that bill STT + LLM + TTS + platform separately. Your finance team should still model tool-call costs, MCP/backend API usage, and human transfer time β the $0.05/min line is the voice path, not the whole COGS story.
For comparison, OpenAI GPT-Realtime-2 uses per-token audio pricing ($32/M input audio tokens, $64/M output, as of its May 2026 launch) plus whatever telephony vendor you add yourself.
| Dimension | xAI Voice Agent Builder | OpenAI Realtime API (GPT-Realtime-2) | DIY STT + LLM + TTS |
|---|---|---|---|
| Primary buyer | Operators, support leads, light devs | Developers building custom clients | Platform teams with integration budget |
| Telephony | Built-in (+ free number) | Bring your own | Bring your own |
| Setup | No-code console | Code + WebSocket | Three integrations minimum |
| MCP | First-class in console | Via your orchestration layer | Via your orchestration layer |
| Pricing shape | $/min flat on audio | Per-token audio + your infra | Three vendor meters |
| Model path | Speech-to-speech Grok Voice | Speech-to-speech GPT-Realtime-2 | Often three-hop latency |
Neither replaces the other automatically: teams already deep on OpenAI Realtime keep their client investments; teams drowning in Twilio + Whisper + GPT + ElevenLabs glue may prefer xAI's bundle.
From the launch thread and replies β the questions that actually showed up:
"How do I get the free phone number?"
Sign up at console.x.ai/voice/agents, create an agent, and provision a number in-console. xAI states every account includes one; exact UI steps may shift during beta.
"Can it navigate phone trees (IVR)?"
Not highlighted in the launch post. Assume outbound/inbound agent flows first; IVR traversal is a verify-in-beta item if your use case is "press 2 for billing."
"Can it call my Hermes agent / sub tools via API?"
Custom API tools and MCP are supported β if Hermes exposes HTTP or MCP, wiring is plausible. There is no Hermes-specific connector in the announcement; plan an integration test.
"Personal assistant for unknown callers β email me details"
Early user reports (e.g. capture name, phone, purpose β email) match the tool + notification pattern Voice Agent Builder describes. Good first-hour beta project; add guardrails before exposing a personal line publicly.
"Is this just a wrapper?"
One reply on X: "functionality wrapper with smart defaults." Fair framing β the differentiation is Grok Voice quality + integrated telephony + MCP at $/min, not a novel agent paradigm. Compare to building on raw GPT-Realtime-2 where you own the wrapper.
Voice is splitting into three lanes on explainx.ai's read:
If your team lives in Claude Code + MCP for coding, a Grok voice line for phone support is complementary β not a migration. Browse live MCP servers on explainx.ai/mcp-servers and skills on /skills for the coding side; use Voice Agent Builder when the channel is PSTN or browser voice, not IDE chat.
Before pointing production traffic at a beta console:
For workshop-style "use Claude as a work partner" without telephony, explainx.ai runs Claude for Work live sessions β different channel, same "AI as colleague" goal.
Build one agent on your hardest workflow β refund line, booking change, tier-1 triage β and call it. Voice quality does not survive benchmark tables alone.
Pricing, benchmark figures, and beta feature availability are accurate as of July 2, 2026 per xAI's announcement. Verify current rates in the xAI console before budgeting production call volume.
