MCP for Hotmart data: ask about sales, refunds and subscriptions without exporting anything
Hotmart has a report, an export and an API. What it does not have is a way for you to ask “how much did I sell of each offer this week, excluding refunds and renewals?” and get the right answer in ten seconds. With an MCP connector on top of the synced data, it does.
Anyone selling on Hotmart knows the ritual: open the dashboard, filter the period, export the CSV, open the spreadsheet, remove the refund rows, separate renewals from new sales, add it up. Half an hour to answer a question the same person will ask again tomorrow. This article is about replacing the ritual with a question.
The path is the Model Context Protocol: a connector that exposes Hotmart data — already synced, already clean — to an AI such as Claude or ChatGPT, with a scope you define. The AI does not get your Hotmart password. It gets a menu of questions it may ask your project.
What Hotmart delivers, and what the AI needs
The Hotmart API returns transactions, and each one comes with product, offer, status, amount, buyer, payment method, installments, and a tracking field (the sck) that carries the UTMs. It is good data. The problem is that it arrives as is: the approved sale, the printed boleto, the expired PIX and the refund are rows of the same report, and a subscription renewal looks exactly like a first purchase.
- Status: only APPROVED and COMPLETE are sales. The rest is intent, delay or return.
- Recurrence: a subscription produces one transaction per cycle. Without separating them, renewals become new sales and the customer count inflates.
- Currency: a sale in dollars or pesos, summed as reais, distorts the whole month.
- Commission: what reaches your account is not the price; it is the producer’s commission after the fee.
A well-built connector solves this BEFORE the question: it syncs transactions into a table per account, keeps history, and documents for the AI what each column means. Then “how much did I sell this week” has one answer.
Questions that become answerable
- 01How much did I sell of each product in the last 30 days, net commission, new sales only?
- 02How many active subscriptions do I have today, and how many missed an installment this month?
- 03Which offer has the highest refund rate in the last 90 days?
- 04How many generated boletos turned into payments, and in how many days on average?
- 05Which campaign (by sck) brought the highest-ticket sales?
- 06Who bought product A and never bought product B?
The sixth question is the one that changes the game, and it only exists because the buyer became a PERSON in the CDP, not a row in the export. Crossing products requires identity, and identity requires leads to be centralized in one place.
Exporting a CSV answers one question. Connecting the AI to the data answers all the next ones.
What the AI must not do
Reading is one thing; writing is another. A sales-data connector should be read-only by default — and “by default” has to be enforced on the server, not promised in documentation. It also needs to refuse queries that cite another project, and return the cost of each question, because a curious AI can scan an entire table without noticing. We have explained what the AI can do on its own and how much a question to the data costs.
How it looks in practice
In CrazyLeads the Hotmart source syncs sales, commissions, subscriptions and participants per account, and the MCP connector exposes all of it to Claude, ChatGPT or Cursor with the project locked in the token. The question “acquisition sales per product, net, this month” becomes a query the AI itself writes, tests and explains. And if the answer deserves follow-up, the AI publishes a report that updates itself instead of a spreadsheet that ages.