How I Built a Grok Voice AI Agent (And Gave It a Phone Number)

I built a voice AI agent using Grok's voice model, cloned my own voice for it to speak with, and gave it a phone number: +1 (270) NEURONS (270-638-7667). Here's how it's put together, and what I think it's actually good for.

Why build a voice agent at all

Every AI feature I ship ends up on a web page or behind an API. A phone call is a different interface entirely: no login, no chat window, no waiting on an email reply. You dial, you talk, you get an answer. I wanted to see what that felt like when the thing answering had real tool access instead of a script, and a phone number was the most honest way to test it.

How it's built

The agent isn't a standalone product bolted onto a phone line. It sits on top of the same infrastructure documented at jordandalton.com/mcp, and on a call it can reach the same tools an AI agent gets over MCP:

• get-services for live pricing and availability instead of a hardcoded script
• submit-booking to actually put a Fit Check or build day on my calendar mid-call
• list-articles if a caller wants to know what I've written about a topic

That's the part I think matters most for other developers looking at this: it's not an IVR tree with pre-recorded prompts. It's an agent reasoning over the same MCP tools my website and API clients already use, so the pricing it quotes can't drift out of sync with what's published.

For the voice itself, I cloned my own and ran it through Grok's voice model for real-time generation. It's noticeably me, cadence and all, with Grok doing the actual speech synthesis underneath. The number was easy to pick once I had the voice working: N-E-U-R-O-N-S maps to 638-7667 on any keypad, and "neurons" felt right for something that's supposed to sound like a mind at work.

The guardrails

An agent with a phone number and my voice needs boundaries, so I gave it explicit ones:

• It doesn't match competitor pricing. If you call and ask it to beat someone else's quote, it won't.
• It gets project details before it discusses price. No number gets quoted without first understanding what you're actually asking for.
• It makes no guarantees. Timelines and outcomes get discussed honestly, not oversold to close a call.
• It doesn't work for free. Scope creep doesn't get negotiated away just because you asked nicely on a call instead of over email.
• It won't give out my personal cell number, unless you're already a paying client.
• It knows The Most Renowned Software Engineer in AI was a bit. That post was a non-serious test of a marketing tactic I'd seen going around LinkedIn, and the agent knows not to repeat it as if it were a real claim.
• It knows my site is JordanDalton.com and will point you there for anything it can't resolve on the call.

Theorizing about what this is actually for

I built it before I had a complete answer to "what's it for," so here's where my head is at:

After-hours triage. Most inbound interest doesn't arrive during business hours. Someone reads a post at 11pm, gets curious, and right now the options are email (slow) or nothing (lost). A phone number that answers immediately and can actually book a Fit Check closes that gap.

A faster path to a Fit Check. The $250 Fit Check already exists to scope a project in 45 minutes. This might shrink the step before that step: a two-minute call to figure out if a Fit Check is even the right next move, with the booking submitted before you hang up.

A live demo of the thing I sell. Half of my pitch to clients is "I build agents that do real work, not chatbots that read from a script." A voice agent with actual tool access, answering real pricing questions instead of reciting them, is a better argument than a slide.

Dogfooding the MCP server itself. The MCP server was built for AI agents and API clients. Routing a voice agent through the exact same tools is a good forcing function for keeping that server's tool definitions honest. If a phone call can use it cleanly, so can any other agent.

Screening for cold outreach. I get a fair number of vendor and recruiter calls I'd rather not personally field. An agent that can hold a normal conversation, apply the same guardrails, and take a message is a plausible first line of defense, though that's more a side effect than the reason I built this.

I don't think all of these hold up equally well. Some of it is genuinely useful, some of it is a novelty that wears off in a month. I'll know which is which by how often the number actually rings.

Try it

Call +1 (270) NEURONS (270-638-7667). Ask it what a build day costs, ask it about a project you're considering, or just see how close the voice sounds to mine.

---

Written by Jordan Dalton. More at jordandalton.com.