Artificial intelligence · 8 min read
MCP server: the standard plugbetween AI and your tools.
MCP (Model Context Protocol) lets an AI such as Claude or ChatGPT use your software: calendar, invoicing, customer records, documents. Here is what it is, what it is for and how to get started, explained without jargon. Start with the video: one minute is enough to get the idea.
In short
- On its own, an AI only knows what it has learned: it can't see your invoices, your customers or your calendar.
- MCP is an open standard that acts as a “plug” between AI and your tools: each tool is connected once, through an MCP server, and can then be used by any compatible AI.
- You decide what the AI can read or do, and sensitive actions wait for your approval.
MCP explained in one minute
Six steps, no sound: all the text is on screen. It follows one simple question (“Has Mr Martin paid his invoice?”), before and after MCP.
Read the transcript
- An AI can answer a lot of questions. Writing, summarizing, translating: it relies on what it has learned.
- But it can't see your tools. Your invoices, your customers, your calendar: all invisible to it.
- Before: a custom cable for each tool. Slow to build, costly, fragile… and redone for every AI.
- MCP: a standard plug, like USB-C. Each tool is plugged in once. Any compatible AI can use it.
- Now the AI can look things up and act. It queries the right tool, then offers to act.
- And you stay in control. The AI only reaches what you open, and you approve sensitive actions.
A brilliant AI, cut off from your tools
Ask Claude or ChatGPT to write a welcome email: the result is often excellent. Now ask it whether Mr Martin has paid his invoice: it has no way of knowing. An AI model only knows what it learned during training and what you paste into the conversation.
In a business, though, the useful information lives elsewhere: in the invoicing software, the customer database (CRM), the shared calendar, the document server, the ERP. Copying and pasting that information into a chat by hand is slow, error-prone and quickly becomes unmanageable.
The traditional answer was to build a connector for every “AI + tool” pair. Three tools and two AIs? Six integrations, each with its own format, to maintain with every update. That is the “custom cable” in the video.
MCP, the standard plug between AI and your software
MCP, short for Model Context Protocol, is an open standard introduced by Anthropic, the company behind Claude, in November 2024. It works like USB-C: one plug shape to connect everything.
- On one side, AI applications that “speak” MCP: Claude, ChatGPT, coding assistants, or your own application.
- On the other, your tools, each exposed by an MCP server.
A tool connected once can be used by any compatible AI. And if you switch AI providers tomorrow, your connections stay in place.
The standard caught on fast: in 2025, OpenAI, Google and Microsoft announced support for it in their products, and at the end of 2025 Anthropic handed the protocol over to the Agentic AI Foundation, under the Linux Foundation, so that it stays neutral and open. In plain terms, it is not one vendor's technology: it is becoming the norm.
What exactly is an MCP server?
An MCP server is a small program that sits in front of one of your tools. It tells the AI: “here is what you can do here, and how”. It can offer three kinds of things:
- Tools, meaning actions: find a customer, list unpaid invoices, book an appointment.
- Resources, data to read: a document, a product sheet, a catalog.
- Prompt templates ready to use: “summarize a customer's file”, “prepare the weekly update”.
Each item is described in plain language. That description is what the AI reads to pick the right tool at the right time.
What happens when you ask a question
- You ask: “Has Mr Martin paid his invoice?”
- The AI sees that an “Invoices” tool is available and decides to use it.
- Your AI application, the MCP client, passes the request to the invoices MCP server.
- The server queries your invoicing software, with its own access rights, and sends back the answer.
- The AI writes the answer and suggests a next step: drafting a reminder. If the action is sensitive, it waits for your approval.
Key pointThe AI never touches your database directly. It only goes through the doors the MCP server has provided.
An MCP server can run on the user's own computer, to reach local files for instance, or on a server, to be shared by a whole team over a secure connection.
Examples in a business
A few typical uses, to adapt to your own tools:
- Front desk & adminFind a free slot, move an appointment, send the confirmation to the customer.
- AccountingList unpaid invoices, draft reminders, track down a payment.
- SalesOpen the customer record before a call, summarize recent exchanges, create an opportunity in the CRM.
- SupportFind the right procedure in the internal documentation, open a ticket, follow up on an order.
- ManagementQuery your data in plain language: “how many orders this month, compared with last month?”
Many vendors, such as GitHub, Notion, Stripe and Atlassian, already provide their own MCP server: for those tools, it is often enough to connect it and configure it properly. Custom work becomes useful for your business software, your ERP or your internal databases.
Security: the AI only reaches what you open
Connecting an AI to your software is worrying, and rightly so. MCP provides a framework, but it doesn't solve security on its own: everything depends on how the server is designed. These are the rules I follow:
- Only what is needed. A server only exposes the actions the use case requires. No “delete” if nobody needs it.
- The user's rights, no more. The AI acts on behalf of the signed-in person, with authentication (OAuth for remote servers), and sees no more than they do.
- Human approval for sensitive actions: sending an email, changing an amount, deleting data.
- A record of every call: who asked for what, when, and with what result.
- Caution with content. A document or an email can contain booby-trapped instructions: this is called prompt injection. So we limit what the AI can do with it, and we don't install an MCP server found at random without checking it.
- Data stays where you decide. The model can be called through a cloud service (Amazon Bedrock, Anthropic API) or run locally, on your own server. The choice depends on how sensitive the data is.
MCP, API, RAG, agent: who does what?
| Term | In one sentence | Where MCP fits in |
|---|---|---|
| API | The technical entry point of a piece of software, designed for developers. | An MCP server often builds on the tool's existing API and makes it usable by an AI. |
| RAG | A technique that lets the AI find the right passage in your documents before answering. | Complementary: a document search can be offered as an MCP tool. |
| Agent | An AI that chains several steps on its own: read, decide, act, check. | MCP gives it its “hands”: the tools it is allowed to use. |
| Connector | An integration built for a single AI application. | MCP is the standard that avoids rebuilding the connector for every AI. |
Does your business need an MCP server?
Probably yes, if:
- your teams often copy and paste information between a piece of software and an AI;
- the same questions keep coming back, and the answer is in your tools;
- you want an AI that acts (schedules, creates, follows up), not just one that writes.
Not yet, if you don't have a specific use case yet, or if your data is scattered across files with no central tool. In that case, start by structuring it.
Where to start
- One specific, measurable use case. For example: “answer questions about order status without opening three applications”.
- The list of data and actions required, and only those.
- A prototype: an MCP server connected to your real data, tested by a few users.
- Going to production: access rights, approvals, logs, monitoring. Then widen to other uses.
Lessons from the field
On my current mission, I build MCP servers and clients connected to cloud AI services (Amazon Bedrock, Anthropic API). A few principles I apply every time:
- A tool's description matters as much as its code. It is what the AI reads to decide whether to use it: a vague name, and it picks the wrong tool.
- A few well-scoped tools beat a whole API exposed as is.
- Short, structured responses. Sending thousands of rows back to the AI slows it down and makes it less accurate.
- Test with real user questions, not just technical tests.
To go further, the Artificial intelligence page presents the building blocks I work with (MCP, RAG, OCR, agents) and lets you put your project together in a few clicks. The official protocol documentation is at modelcontextprotocol.io.