Connecting Claude or ChatGPT to your sales calls with MCP

You have probably pasted a call transcript into ChatGPT or Claude to get a summary or draft a follow up. It works, for one call. By the tenth, the approach shows its limits: the conversation overflows, the assistant has forgotten the first call, and confidential words from your customers are travelling through copy and paste.

The Model Context Protocol, or MCP, takes another route. Instead of bringing your calls to the assistant, you give the assistant read access to the place where they are stored. Here is what that changes for a sales team, what you can ask, and the precautions to take before plugging it in.

MCP in plain words

The Model Context Protocol is an open standard introduced by Anthropic in November 2024. It describes a common way for an AI assistant to call tools and read data held in other applications. An application that adopts it exposes an MCP server: a list of tools, each with a name, a description of what it does and the parameters it accepts.

When you ask a question, the assistant picks the relevant tools itself, calls them, reads what they return and writes its answer from what it retrieved. Claude, ChatGPT, Claude Code and Cursor can all connect to an MCP server, which spares you from building a different integration for each of them.

Pasting a transcript lends the assistant a document. Connecting it through MCP opens the library to it, read only.

Why pasting transcripts does not scale

Copy and paste helps now and then. It runs into four limits as soon as you want to reason about a book of accounts rather than a single call.

• The context window. A model only reads a limited amount of text at once. One hour of call fits, twenty hours of exchanges with one account far less well, and all of your customers never.
• No memory from one conversation to the next. Every new chat starts from zero. To compare two accounts or follow a deal over three months, you paste everything again, every time.
• Confidential data on the move. Transcripts hold names, budgets, internal tensions. Copying them by hand into personal chats multiplies the copies, with no trace of who shared what.
• No verifiable citation. The assistant summarises what you gave it, but nothing ties its sentence to the exact passage of the call. To check, you have to reread the whole transcript.

Add a quieter bias: you paste the calls you remember. The analysis therefore covers a selection made from memory, rarely everything your customers actually said.

What a connector to your customer conversations changes

With a connector, the assistant no longer receives a text, it receives tools. Linked to a base of customer conversations already transcribed and analysed, it can for instance:

• search your exchanges using the words a customer would have used, filtered by customer, period, exchange type or deal stage;
• read a customer file: the account synthesis, its deals, the commitments still open and the latest exchanges;
• count themes: group the passages of one category, objections for example, and say how often each subject comes back instead of guessing;
• quote passages: back every claim with an excerpt, saying who said it and in which exchange, and read the neighbouring passages to put a sentence back in context.

The difference fits in one sentence: the assistant no longer works on what you thought of giving it, but on everything the team captured, and it fetches what it needs to answer on its own.

Ten questions to ask your assistant

Here are concrete requests a sales leader can make once the connector is in place. Names and situations are examples; the quality of each answer depends on what your exchanges actually contain.

Prepare and follow an account

• "I am meeting Axelia on Thursday. Summarise the relationship, the commitments still open on our side and the questions left unanswered on the last call."
• "What has the customer promised to do since our last demo, and what did we promise them?"
• "Recap this account in ten lines for the person taking it over, with a quote for each important point."

Understand what keeps coming back

• "Which objections come up most in our demos over the last three months? Give the number of passages for each."
• "Which competitors are mentioned during negotiations, and about what?"
• "What do customers say when they talk about budget? Quote three representative passages."

Compare and learn

• "Compare what customers said on won deals and on lost deals: which arguments come up on one side and not the other?"
• "For the finance director persona, which themes come back most often?"
• "Which customers expressed a churn risk in recent weeks, and in what words?"

Prepare an action

• "Draft a follow up for this customer using the exact sentence where they describe their problem. I will review it before sending."

That last example needs a word of caution. A message written by the assistant stays a draft: a person reviews it, checks the passage it quotes, adjusts it and sends it under their own name. The connector reads, it sends nothing.

Security and governance: the checklist

Opening customer conversations to an AI assistant is not a small step. Before plugging in a connector, whatever the tool, check the following points.

• Read only. The connector must be able to read without being able to modify, delete or trigger any paid action.
• A limited scope. The assistant must only see the workspaces the connected user already belongs to.
• Explicit consent. The connection goes through an authorisation screen showing which client is being granted access, never through a shared password.
• Easy revocation. You must be able to cut an assistant's access at any time, and see which ones are really connected.
• Who sees what. Each person who connects their assistant only reaches what they already see in the application. Decide who in the team needs this use.
• The model provider. The passages the assistant reads are sent to the provider of that model. Check its terms and your account settings, in particular on whether data is used for training.
• GDPR and informing the people recorded. A recorded call holds personal data. Participants must be informed of the recording and its use, and the purpose must remain knowing your customers, not rating individuals.

A good connector creates no new right: it opens to the assistant what the user already sees, no more, no less.

Connecting Claude or ChatGPT to Meidly

With us, the connector is part of the product. Your calls, recorded on Google Meet with the Chrome extension or imported as an audio file, a video file or a link to a recording, are transcribed, attached to a customer file and cut into tagged passages. That base is what the connector opens to your assistant.

Three steps

• Open the MCP, API & Integrations page of your account.
• For Claude or ChatGPT, add the connector from your assistant, then authorise access in one click on the consent screen. No token to copy.
• For other MCP clients, such as Claude Code or Cursor, generate a token from the same page and add it to the client's configuration.

The connector is read only: it writes nothing to your data and consumes no credit. It gives access to the workspaces, the customers, the exchanges and their passages, the counted themes and the personas. The page shows which assistants are actually connected, and you can disconnect any of them at any time.

The limits, honestly

A connector does not make the assistant infallible. Three limits are worth keeping in mind.

• The answer is only as good as what was captured. A call that was not recorded or a decision made in a corridor does not exist for the assistant. If it finds nothing, that does not prove nothing was said.
• A model can misread. It can take a sentence out of context, mix up two speakers or generalise from two passages. A transcript can also contain recognition errors.
• The citation is there to be checked. Before reusing a claim in a decision or a message, open the quoted passage and read what surrounds it.

Used this way, the assistant does not replace review, it makes review possible: instead of listening to twenty calls again, you check the three passages its answer rests on.

Where to start

• Gather the recordings of a few active accounts, with the participants' consent.
• Connect a single assistant, for one or two people in the team, and look at what it sees.
• Start with questions you already know the answer to, to judge the quality of the citations.
• Then widen to questions of frequency and comparison, where the gain is clearest.

The benefit does not show on the first question. It shows the day you ask what keeps coming back across fifty calls and get a counted answer, with the passages to check it.