MCP: how to connect Claude or ChatGPT to your restaurant's data
MCP is an open protocol through which external AI reaches data by a set of read tools rather than by file exports. The owner approves the connection, the AI requests the slices it needs itself, every call is logged, nothing is written back into the restaurant's systems, and access is revoked in one action.
The usual way to show AI your figures is to export a spreadsheet and attach it to the question. It works exactly once: the next question needs a new export, and the data in the old one is already stale.
MCP (Model Context Protocol) does it differently. Instead of a file, the AI gets a set of tools and requests the data itself, at the moment it needs it to answer.
What it looks like from the restaurant's side
You add the connection address in Claude, ChatGPT or any other client that supports MCP.
The client opens the RestoAudit sign-in page - you approve the access with your own account.
Then you simply ask: «why did food cost rise in August», «which suppliers put prices up», «what is going on with bar stock».
The AI decides which slices of data it needs, fetches them and answers in words.
What exactly is read
Purchasing and invoices, price history, dish cost and recipe cards, sales, stock, write-offs, suppliers, shifts, signals and reports. The same data you see in the app - not one line more.
The limits that cannot be crossed
Read-only. There is no write path into the restaurant's systems: the AI creates no documents, changes no prices, posts no invoices.
The owner approves. Without an explicit approval the connection does not work.
A log. Every call is visible: who, when, which data.
Revoked in one action. Access is removed in settings, with no email to support.
No other company's data. The connection sees only the companies you have access to.
Detail and limits are on the external AI connection page and in the security section.
How this differs from a chat with an export
The data is fresh: it is read when the question is asked, not when the export was made.
You do not have to guess in advance which slice will be needed.
The question can be refined - the AI fetches whatever it was missing.
Nothing is stored in the conversation: a file of prices does not move into a chat for good.
Who this is for
An owner of several venues who wants an answer rather than a dashboard. A manager testing one hypothesis who does not want to open five tabs. An accountant reconciling a period. And a partner building their own solution on top of our data - for them it is the same protocol.
External AI connectionSecurityDoes the restaurant's data go into model training?
We do not train models on your data: under the data processing agreement it is used only to provide the service. What your AI client does with the conversation is set by its own settings.
Can the AI change anything in the books?
No. The connection is read-only - there is no write path into the restaurant's systems.
What is visible in the log?
Who made the call, when, and which data was requested. The log is available to the owner in settings.
How is access switched off?
In one action in the connection settings. Requests stop going through immediately.