Cuedis

Letting Claude or ChatGPT use your business data safely (MCP explained simply)

AI assistants like Claude, ChatGPT and the coding tool Cursor are good at writing and reasoning. On their own, though, they know nothing about your business. Ask Claude which supplier has been late this month and it can only guess.

MCP, the Model Context Protocol, is the standard way to fix that. It lets an AI assistant use tools and data from other software, including your business systems. This post explains what it is in plain terms, and how to set it up so the AI gets useful answers without getting your whole database.

What MCP is

MCP is an open standard, first published by Anthropic in late 2024, that describes how an AI assistant can talk to other programs. It is now supported by many assistants and tools, including Claude, ChatGPT and Cursor.

Think of it as a plug socket for AI:

  • A program that has useful information, such as a CRM, a file store or a business database tool, offers an MCP server. It says: "Here are the things I can do: list your suppliers, look up a customer, show this week's concerns."
  • The AI assistant connects to it as an MCP client. When you ask a question, the assistant decides which of those things to use, asks for them, and uses the results in its answer.

So when you ask "Which supplier should I worry about?", the assistant can call a tool that returns the actual concerns from your business, instead of making something up.

Why not just connect the AI straight to the database?

It is possible to give an AI assistant a database login and let it write its own SQL. For anything other than experiments, it is a bad idea:

  • It sees everything the login can see, including personal data and figures that most people in the company should not see.
  • Write access is dangerous. A login that can change data, used by a tool that writes its own queries, can change data by mistake.
  • No business knowledge. The AI has to work out from table names alone what everything means. Old POS systems with columns like TRNTYP and DEPT5 do not make that easy, and the answers can be confidently wrong.
  • No record of who asked what.

A well-designed MCP server sits in between. It offers a limited set of well-defined tools, enforces who may see what, and returns answers that already make business sense.

What to check before connecting business data

If you connect any MCP server to your business data, check these points:

  1. Read-only. The server should only read. Ideally its own database login is read-only too (see our guides for SQL Server and MySQL).
  2. A separate key per connection. Each person or laptop that connects should have its own key, so you can see who is connected and remove one without affecting the others.
  3. Limits per connection. A connection for the purchasing team should see suppliers and stock, not salaries or customer phone numbers.
  4. Personal data off by default. Unless there is a real need, personal details should not be available to the AI at all.
  5. On your network, not the internet. If the server runs in your office, it should only be reachable from your office network.
  6. A record of what was asked. You should be able to see which connection asked what, and when.
  7. It remembers who changed what. Turning on a connection or adding a key is an important change. The owner should be told.

Where the AI runs matters too

Remember that when you use Claude or ChatGPT, your question and the answers from the MCP server go to that AI provider. A limited MCP server keeps that to the facts needed for each answer. But it is still worth deciding which information should be available through it. (We wrote more about this in AI on your own PC or in the cloud.)

How Cuedis does it

Cuedis includes an MCP server, so AI assistants on your office laptops can use it as their business memory. The tools are things like list businesses, concerns, briefing, facts, history, ask and add a note. They answer from what Cuedis has already learned, not by letting the AI loose on the database.

Each connection gets its own key and is tied to a team, so it can only see what that team may see. Cuedis only reads, personal details stay off unless you switch them on for that business, and the MCP server is only reachable on your office network. Turning AI agents on, and every new connection, is reported to the admin and the owners. It is part of the paid licence and is switched on per business under Settings → AI agents, which also shows the configuration to paste into Claude Desktop or other tools.

MCP is a genuinely useful step: it lets the AI assistant you already use answer questions about your actual business. Just make sure that what you plug in is read-only, limited, and on your terms.