Blog · May 24, 2026 · 5 min read

How to Build a Voice AI Agent for a Small Business

Give a Hermes agent a real phone number with AgentCall, write a receptionist prompt, and it answers calls for a business. Full setup and a live test call.

Watch the video: Hermes + AgentCall: Build Voice AI Agents for Small Businesses

You build a voice AI agent for a small business by giving a Hermes agent a phone number through AgentCall, then telling it who it works for and what to collect. Get an API key, add it to the server running Hermes, ask the agent to install AgentCall, provision a local number, and paste in a receptionist prompt. In this guide the agent becomes the after-hours receptionist for a 24-hour plumbing company, and I call it live.

The test call is worth hearing, so watch the full walkthrough. AgentCall is my own product; I built it because agents needed a phone layer that was simple. Here is the setup.

What AgentCall gives the agent

AgentCall is inbound calls, outbound calls, voice and memory in one API key. The free tier gives you one US local number, five voice minutes, five inbound AI voice minutes, and a post-call summary if your Hermes is connected to email. The Pro plan is $19.99 a month plus a few cents per use, and each additional number is $2 a month.

For the agent side, AgentCall ships as an MCP server with about 45 tools, so once it is installed the agent can provision numbers, configure them and place calls on its own.

Step 1: create an API key

  1. Sign in at AgentCall with Google or GitHub.
  2. Click API key, then Create key. I named mine "Hermes Agent."
  3. Copy the key.

Step 2: add the key to your server

My Hermes agent runs on a Hostinger VPS.

  1. Open the Hermes instance and click Manage.
  2. Scroll to Environment and add a variable named AGENTCALL_API_KEY with the key as its value.
  3. Click Save and deploy.

Step 3: have the agent install AgentCall

In Telegram, tell the agent: "Please install AgentCall using the MCP. I have added the AgentCall API key into the environment."

The agent reads the MCP server, pulls in the tools and documentation, and installs. It takes a couple of minutes. For this instance I was running Hermes on a ChatGPT subscription with GPT-5.5, which handled it fine.

Step 4: provision a phone number

Ask for a number in your area: "Configure AgentCall fully. Provision a local area code number for a receptionist." I asked for a 636 area code, which is local to me. If you want a toll-free number, ask for an 800 number instead.

The agent read the pricing first and checked with me that a number carries a $2 monthly charge before creating it. I confirmed. A few seconds later it returned the number, its ID, the type, the area code and the rate, and marked it active. You can confirm the number exists in the AgentCall dashboard.

Step 5: give the receptionist its context

A number alone is not a receptionist. The agent needs to know who it works for. Here is the prompt I used for a 24-hour plumbing business:

"You are a receptionist for Joe's Plumbing, a 24-hour service business that offers emergency services for all homes and businesses in the St. Louis area. You are responsible for getting clients' information, including name and address, type of home or business, the level of service they are looking for, and how quickly: today, tomorrow, next week. Be kind and respectful, and only focus on getting their information to route to our owner, Joe Johnson, who will call or text them back at his earliest availability."

If you have never written a prompt, ask the agent to draft one from the client's needs, or use ChatGPT.

The agent came back with a summary to confirm before it enabled autonomous answering: the number, business name, service area, availability, its job, the owner's name, recording off, no email summaries yet, and a suggested five-minute cap per call to prevent runaway calls, with the budget that implies. That check is exactly what you want before a number starts answering on its own.

I added one thing: the greeting. "When answering the call, say: 'Joe's Plumbing. I'm Joe's virtual assistant. How may I help you today?'" More context produces a better receptionist. Then I confirmed and asked if we were ready to test.

Step 6: the test call

I called the number on speaker. The receptionist answered with the greeting, and then, when I said I had an emergency with a toilet, it:

  1. Expressed concern and asked for my full name and the best callback number.
  2. Read the number back and asked me to confirm.
  3. Asked for the service address, and confirmed it.
  4. Asked whether it was a home or a business.
  5. Asked how quickly I needed service.
  6. Asked for any access notes for the plumber.
  7. Summarized everything it had collected and said it would send the details to Joe, who would call or text back at his earliest availability.

That is a complete intake call, handled without a person, and Joe gets the summary.

Running this for more than one client

If you set up receptionists for several businesses, do not rely on the agent's memory alone to keep their details apart. Give it a database to store each client's information. Databases come with MCPs now, so you can ask the agent which one to use. Turn on recording and add an email for summaries per client as needed.

What this means for your business

  • After-hours calls answered. A solo tradesperson on a job cannot pick up. The agent does, collects everything, and sends a summary.
  • No missed emergency leads. The caller with a broken toilet at midnight gets a calm, complete intake instead of voicemail.
  • Multilingual by default. The same agent handles callers in other languages; I demonstrate that in the use cases video.
  • A cap on cost. The per-call minute limit keeps the bill predictable.

The phone is one of the six use cases I run, and it pairs well with Gmail access so summaries go straight to the owner's inbox. If you would like a receptionist number set up for your business, get in touch and tell me what your callers usually need.

FAQ

What is AgentCall?

A platform that gives an AI agent inbound and outbound calling, voice and memory through one API key. It installs into Hermes as an MCP server with around 45 tools.

How much does a voice AI agent cost with AgentCall?

The free tier includes one US local number and five voice minutes. The Pro plan is $19.99 a month plus usage, and each extra number is $2 a month.

How does the agent know what to say on a call?

You give it a prompt describing the business, what information to collect, and how to greet callers. The agent confirms the configuration with you before it starts answering.

Can I stop calls from running too long?

Yes. The agent suggested a five-minute cap per call to prevent runaway calls, and you can set your own limit.

Can one agent handle several businesses' phone lines?

Yes, but give it a database to keep each client's details separate rather than relying on memory alone, and provision a number per business.

More guides · Video tutorials · Work with Percy