What is MCP? The Model Context Protocol and MCP servers explained

MCP in one paragraph

The Model Context Protocol (MCP) is an open standard for how an AI assistant talks to other software. Without MCP, every AI app needs its own integration for every program. With MCP, a software vendor builds one MCP server, and every AI app that supports MCP can use it. That is why MCP is often compared to a USB-C port for AI: one plug for everything.

Anthropic, the maker of Claude, published MCP in November 2024. In December 2025 Anthropic handed the protocol to the Agentic AI Foundation, a fund under the Linux Foundation it co-founded with OpenAI and Block, among others. The specification has since been developed in the open at modelcontextprotocol.io.

What is an MCP server?

An MCP server is the program that offers a connection. It tells the AI assistant what is possible and carries out its requests. There are three roles:

A server can offer three kinds of things:

• Tools: actions the language model can take, such as "find open invoices" or "create a dashboard". The model decides when to use a tool; the app asks for your permission.
• Resources: data that can serve as context, such as a document, a table or a schema.
• Prompts: ready-made instructions you pick as a user, for example "build a monthly report".

The app can offer things to the server as well. With elicitation a server asks the user for more information, such as which company file is meant. With the MCP Apps extension a server can show interactive elements in the conversation, such as a chart or a form.

How does MCP work technically?

For those who want to know: MCP messages are JSON-RPC 2.0 messages. The client sends a request, for example tools/list to see which tools exist, or tools/call to run one, and the server returns the result. In the July 2026 version the protocol is stateless: every request carries its own protocol version and client capabilities, so a server does not need to remember what happened earlier on the same connection.

There are two ways to connect:

For business data a remote server is almost always the better choice. Nothing to install, no passwords or API keys shared with the AI app, and the vendor can grant and revoke access per user.

Which AI apps support MCP?

MCP became the standard quickly. These apps, among others, can use MCP servers:

• ChatGPT: through apps and, in developer mode, as a custom MCP connector. See connect ChatGPT to your business data.
• Claude: through connectors in Claude.ai and the desktop app. See connect Claude to your business data.
• Microsoft Copilot Studio: for agents within Microsoft 365.
• Code editors such as Visual Studio Code with GitHub Copilot, and Cursor.

Which features are available differs per app and plan. On business plans an admin often decides which connectors staff may use.

Which business software has an MCP server?

More and more vendors offer their own MCP server. For accounting software the picture is mixed: Moneybird has its own server, Exact Online and Twinfield do not, though third parties offer connections. See connect Exact Online to ChatGPT or Claude and Twinfield with Claude via read-only MCP.

MCP, API and OData: what is the difference?

Behind the scenes an MCP server often simply uses the software's API. The difference is who drives it: with MCP it is the language model. When to choose OData and when MCP is covered in OData and MCP: what's the difference?.

Is MCP safe?

The protocol handles transport and sign-in; how safe a connection is depends mostly on the server and on what you allow. Watch for these risks:

• Too many permissions. A server that can write, delete or pay can do so when the model misreads a question. Prefer read-only where possible.
• Prompt injection. Text in a document, email or web page can try to make the assistant do something else. The less a server may do, the smaller the damage.
• Unknown servers. A tool description can be misleading. Use servers from the vendor itself or from a party you trust.
• Data at the AI provider. What your assistant retrieves ends up in the conversation and falls under the terms of your plan with that provider.

More on agreements and settings in securing AI access to business data and on the page using AI safely on your business data.

Example: the DigiData MCP server

DigiData offers a remote MCP server for ChatGPT, Claude and other apps that support MCP. This is how it works:

1. Connect your software to DigiData, for example Xero, QuickBooks Online or Exact Online. The data is then updated automatically in a secured database in the EU.
2. In ChatGPT or Claude, add the URL https://mcp.digi-data.nl/mcp and sign in to DigiData with OAuth.
3. Choose which data the assistant may see.
4. Ask your question, for example: "Which invoices have been open for more than 60 days?" or "Which projects are running over on hours?"

Your sources stay read-only: the assistant cannot change or send anything in your accounting system, and you share no passwords with ChatGPT or Claude. An admin can optionally let the assistant create and change DigiData dashboards and saved calculations; it then asks for confirmation before every change. Because the data is already prepared, the assistant does not have to call your accounting software for every question. Answers are faster and you do not run into the API limits of your package.

MCP or your own AI agent?

MCP is useful when your team already works in ChatGPT or Claude. Want everyone to use the same definitions, get answers in Teams or Slack and receive reports automatically? Then the DigiData AI agent is the stronger choice. Read why teams choose DigiData AI.

Get started

Follow the guide for ChatGPT or Claude, see what to do when the connection does not work, or request a demo.