MCP servers

What is an MCP server? Think of a plug socket

MCP stands for Model Context Protocol, an open standard that lets an AI model use external tools and data. In plain terms: connecting your management software to a model used to mean making a custom cable for that pair. Change the model and you needed another cable. MCP is the standard socket: you fit the socket once (the server) and any model that understands it plugs in without new building work. That server is a program running on your side. Inside it, tools are defined: check stock, see free slots, find a case file by number. The AI doesn't see your database; it sees that list and can call them.

Source: Public specification of the Model Context Protocol

What exactly does the AI see when it connects?

The AI isn't handed a copy of your database or a folder of your files. It's given one specific tool, with its input and output defined. "Check availability" returns next week's free slots, not the whole diary with every patient's name. "View order" returns the status of the order whose number the customer has given, not everyone else's history. Each tool does one thing, returns only what's needed and logs who called it and when. That's what gets designed in an MCP project, and it's the part you don't see in a demo.

If you change model tomorrow, the MCP stays put

This is the strong argument, and it's worth being clear on before signing anything. The market moves fast: today you use one model and tomorrow there may be another that works better, or that your customer requires under internal policy. If the connection to your data is written for one specific provider, switching means rebuilding it from scratch. With an MCP server, it doesn't: the protocol is public and tools from different vendors speak it. You change the model and what you defined keeps working. You take the socket with you. We apply this in-house: we publish an llms.txt, a plain-text index of the site designed for models, and our robots.txt gives explicit permission to GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot. Open it and check.

Source: The llms.txt we publish on xoraagencia.com

A badly built MCP is an open door to your data

It needs saying plainly, even if it doesn't help sell. An MCP server exposes your company's data to a program that decides on its own what to call. If it's built in a hurry and without care (a tool that runs any query against the database, with no authentication and no limit on what it returns), what you've installed is a hole, and a convenient one at that. We build it the other way round: what's exposed and what isn't is decided before writing a line; each tool asks only for what it needs; what writes is kept separate from what only reads; and every call is logged. Any personal data that goes out is kept to a minimum, as the GDPR requires.

Source: General Data Protection Regulation (GDPR), Article 5(1)(c) (data minimisation)

What can be exposed: the catalogue, the diary, the case files

The catalogue. The tool that checks stock returns what's in your system right now, not what was there at the last export, and if something is out of stock it says so. The diary. One tool reads the free slots and a different one creates the appointment: they're two separate permissions, and you can choose not to grant the second. The case files. Your team asks in everyday language how a client's matter is going, and the query is limited to what that person is entitled to see: someone who can't open a file can't see it by asking either. Who then uses those tools (the WhatsApp assistant, your own team) is a separate decision.

When you don't need an MCP (and what you need instead)

If all you want is for the AI to answer questions about your documents (the manual, the internal procedure), you don't need an MCP. You need a knowledge base: the documents indexed so the model can search them and answer citing where it got it from. It's simpler and solves that problem better. MCP comes in when the answer depends on data that changes (what's in the warehouse today, which slot is left on Thursday) or when you want the AI to do something, not just talk about it. We do both. Tell us about your case and we'll tell you whether we can do it, how we'd approach it and what it would cost. If it isn't for us, we'll tell you that too.