Connecting Claude, ChatGPT and Cursor to your call data via MCP

What a Claude MCP telephony integration server is

A Claude MCP telephony integration server is a web address that speaks the Model Context Protocol, the open standard AI assistants use to reach outside tools. Connected to an AI phone line, the server gives Claude, ChatGPT or Cursor a fixed list of tools for reading call records, so the assistant answers from your real data.

It helps to know what MCP is not. It is not the AI that answers your phone. That is your AI phone agent, and it carries on answering every call whether or not you ever connect an assistant. MCP is a side door for you, the owner, to ask questions afterwards.

The difference from a dashboard is the shape of the question. A dashboard shows you lists. An assistant can answer something like: which callers this week asked about weekend hours, and did any of them book? It works that out by calling several tools in a row.

What the Prepaid AI server can read and do

The Prepaid AI MCP server gives an assistant read access to one workspace: the plan, the balance, agents, numbers, calls with transcripts, messages, bookings, contacts, free appointment times and the Brain. The server offers one action that changes anything, a booking request, and that request waits for a person's approval.

The full tool list

Two of these deserve a note. check_availability reads from the same booking engine your phone agent books from, so the assistant and the AI on the phone never disagree about what is free. And search_brain returns source wording from your Brain, not an answer composed on the spot, so you can see exactly what your AI would be working from. There is more on that in how an AI phone learns your business.

The one thing it can ask for

book_appointment is the only tool that is not a pure read, and it only asks. The assistant sends a start time and a customer name, gets back an approval id, and stops. The request shows up in your console, where a person approves or denies it. Only then does the booking exist. The assistant can call get_approval to learn the result and tell you.

That is deliberate. An assistant can misread a date, and a diary entry is a promise to a customer, so a person makes it.

What it cannot do

Nothing reachable from an AI session can spend credits, buy a number, publish changes to your agent, take an agent live or delete anything. You can also switch individual tools off in the console, or set them to ask first.

How to connect each assistant

Claude and ChatGPT

Both connect by sign-in. You never paste a credential.

1. In your assistant's settings, add a custom connector. Some versions call it a remote MCP server.
2. Give it the server URL: https://app.prepaid.ai/api/mcp.
3. The assistant opens your browser at prepaid.ai. Sign in if you aren't already, check the name of the app asking for access, and approve.
4. Ask it something. Try asking what urgent messages came in today, or to list your phone numbers.

The connection is to the workspace you are signed in to when you approve.

Behind the scenes this is OAuth with PKCE, and the only scope is workspace read.

Cursor and other clients set up from a file

A client with no browser approval step sends an API key as a bearer token instead.

1. In the console, open Settings, API keys and create a key. It is shown once, so copy it then.
2. In the client's MCP configuration file, add a server entry under mcpServers with the same URL as above.
3. Add an Authorization header to that entry with the value Bearer followed by a space and your key.
4. Restart the client and check the tools appear.

It is the same key the REST API uses, so keep it out of shared repositories and revoke it if it leaks.

Ending a connection

Settings, Connected apps lists every assistant you have connected and when each last used its access. Disconnecting one ends it immediately and everywhere. For a key-based client, revoke the key.

Things worth asking once it is connected

The value shows up in questions that cross several lists. A few that work well:

• Summarise yesterday's calls, and tell me which ones need a call back.
• What did the caller from this morning's urgent message actually say? Read me the transcript.
• Which questions has my AI been unable to answer this month, most common first?
• How many plan minutes do I have left, and when does the plan end?
• What does my Brain say about our cancellation policy?

The unanswered questions one is the quiet winner. Each item on that list is a caller who asked something your AI had no answer for. Ask your assistant to draft answers, then add the good ones to your Brain yourself.

With the server connected in Cursor, developers can ask about real call records while writing the code that handles them. If you want events pushed to your own server instead, see webhook routing for call data.

Is it safe to connect an assistant to call data

Three things limit the exposure. Access is to one workspace only. The tools are fixed, so an assistant cannot go looking through anything outside that list. And the single action that changes something waits for a person.

What the server cannot control is the assistant itself. Once Claude or ChatGPT has read a transcript, that content is in your conversation with that provider, under their terms and your settings there. Connect the assistants you already trust with business information, and disconnect any you stop using.

Every tool call is recorded in the console, so you can see which assistant asked for what.

Try it on your own calls

MCP comes with the Standard and Pro plans. Standard is US$79 for 350 minutes over 30 days, and the pricing page shows what each plan includes. You can start with the 7 day free trial, no card needed, and talk to your AI in the browser before you pay for anything. Other connections are on the Integrations page.

FAQ