What is MCP? And what's in it for your business

MCP (Model Context Protocol) is an open standard that lets an AI assistant such as Claude, ChatGPT or an agent in Copilot Studio work with your own software. An MCP server tells the assistant which actions are available in a system, such as looking up a customer or preparing a page, and carries them out. For a business, that means your team works in plain language in your systems, within boundaries you set in the server.

The short definition is in the glossary entry MCP. Here you'll read what's in it for your business, how it differs from an API and an AI agent, how we use it and what to watch to keep it safe.

What exactly is MCP?

MCP is an agreement on how an AI application and other software talk to each other. Anthropic introduced the standard on 25 November 2024 (Anthropic). On 9 December 2025, MCP moved to the Agentic AI Foundation, a fund under the Linux Foundation (MCP blog). So the standard doesn't belong to one vendor.

MCP has three roles (specification, version 2025-11-25):

• Host: the AI application you work in, such as Claude or ChatGPT.
• Client: the part of the host that keeps the connection to one server open.
• Server: the program that opens up a system, such as your CMS, CRM or accounting.

A server offers three kinds of things: tools (actions the assistant can take), resources (data it can read) and prompts (ready-made templates for a task). The assistant first asks what's available, then picks what fits the request.

What is an MCP server?

An MCP server is the program that makes one system available to an AI assistant. It describes which actions exist and what each one needs, and usually calls that system's API to carry them out. Whatever the server doesn't offer, the assistant can't do either.

Some vendors offer their own MCP server, such as HubSpot and Salesforce. How to use those is covered in connecting Claude to HubSpot and connecting Claude to Salesforce. For Exact Online and AFAS, we didn't find a server from the vendor itself when we checked on 18 October 2026. There, the route is your own server on the package's API, as described in connecting Exact Online to Claude via MCP and connecting AFAS to Claude via MCP.

What's in MCP for your business?

You build one integration per system, and every assistant that supports MCP can use it. Your team asks questions and gives instructions in plain language, and you set in one place what is and isn't allowed.

• Asking your own systems questions: which invoices are open for this customer, who starts next month, which tickets are past their deadline. Without an export, and only with the fields the question needs.
• Preparing work: putting a page in your CMS as a draft, preparing a draft quote or creating a task, which a colleague then reviews.
• Not being tied to one assistant: Claude supports MCP through custom connectors (Claude Help Center), the OpenAI API through the MCP tool in the Responses API (OpenAI) and Copilot Studio through MCP tools and resources (Microsoft). Checked on 27 October 2026.

It isn't a goal in itself. If your team doesn't work in an AI assistant, or a task's steps are fixed, a regular integration serves you better.

How do we use MCP?

With a website, we deliver a custom MCP server so the team can create and edit pages through Claude or ChatGPT using the building blocks of their own design system. Changes are first ready as a draft with a preview.

At Sprints & Sneakers, the team manages the website through Claude this way: pages, blogs, cases and events. Everything is first ready as a draft with a preview, and publishing is a separate step. How that fits together is in managing your CMS with your own MCP server.

That's our starting point with other systems too: start with read access, only add write actions once it's clear who approves what, and put every change up as a draft first.

What's the difference between MCP, an API and an AI agent?

An API is a system's door for software that knows in advance what it needs to do. An MCP server usually sits on top of it and describes the actions so a language model can find and choose them itself. An AI agent is who acts: the system that's given a goal and takes steps itself.

The agent decides what it wants to do, the MCP server decides what it can do. More on agents in what is an AI agent, and on the integrations underneath in connecting an AI agent to your existing systems. The glossary entry API integration explains the basics.

Is MCP secure?

MCP is as secure as the server you connect. The specification asks that a user consents before a tool is called, but adds that the protocol itself can't enforce that. That's down to whoever builds and configures the server. According to the specification, tool descriptions shouldn't be trusted unless the server comes from a trusted source.

• Only the actions you need: if your team may prepare pages, don't build an action to publish or delete.
• Minimal permissions: the server gets only the rights in the underlying system that its actions need. Start with read-only.
• Draft first: whatever the assistant creates or changes waits for a person to review it.
• Logging: record which tool was called when, with which data.
• Watch for misleading input: if the assistant reads outside text, such as an email or web page, instructions can be hidden in it. See prompt injection.

If personal data passes through a server run by another party, that party is a processor and you need a data processing agreement. What else to arrange is in the AI Act and GDPR: what do you need to arrange?.

When don't you need MCP?

When there's no need for a language model in between. If the steps are fixed, such as an order from your webshop that has to reach your accounting, a regular integration is simpler, cheaper and easier to control.

• Your system already has a ready-made app or connector in the assistant your team uses, and it does what's needed.
• You're building AI into your own application and define the tools in your own code. Then MCP is an option, not a requirement.
• Nobody on your team works in an AI assistant. Then start with the work, not the integration.

Build your own MCP server or have it built?

To try it out, you can connect an existing vendor server to Claude or ChatGPT. When it involves business data, write actions or several users with different permissions, you want a server you manage yourself, with logging and a clear owner.

We build such servers as part of having an AI agent built or a website you manage through Claude. Not sure yet which work pays off most, or whether MCP is even the right answer there? The AI Opportunity Audit first maps what the current work costs and where the opportunities are.