What is MCP? The Model Context Protocol in plain words

You ask your AI assistant how last week’s sales compare with the week before. It writes a helpful paragraph about how to find out, because it cannot see your sales. The numbers are in your store, your ad account and your CRM, and the assistant has no way in. So you open each one, copy the figures, paste them into the chat and ask again.

MCP is what removes that step. This post explains what it is in plain words, where it came from and what it does and does not change. If you already know the protocol and want the difference between a server and a connector, we covered that in What are MCP connectors?

The short answer

MCP stands for Model Context Protocol. Its own documentation describes it as an open-source standard for connecting AI applications to external systems: files, databases, search, and the apps you use for work.

The same page offers an analogy that holds up well. MCP is like a USB-C port for AI applications. Before USB-C, devices needed their own cables. Without a shared standard, each AI app has to be connected to each service separately. With one, a service can be plugged into any app that supports it.

Where it came from

Anthropic, the company behind Claude, announced MCP on 25 November 2024. The announcement named the problem directly: even the most capable models are “constrained by their isolation from data”, cut off behind information silos and older systems. Anthropic released the specification, software kits for developers and a set of example servers, all as open source.

A year later, on 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation, a new foundation under the Linux Foundation. That matters for you in one way: the standard is not tied to one AI company. The protocol’s documentation lists support across AI assistants such as Claude and ChatGPT and developer tools such as Visual Studio Code and Cursor.

How it works, in three parts

The architecture has three roles. You only ever see the first one.

• The app you type into. Claude, ChatGPT, a code editor. The specification calls it the host.
• An MCP server. A program that reaches one outside system, such as your store or your analytics, and tells the app what it can do there.
• A client inside the app. The app opens one of these for each server it is connected to. It carries messages back and forth, and you never deal with it directly.

When the app connects to a server, the server hands over a list of what it offers, each item with a name and a short description written for the model. When you ask a question, the model reads that list, decides which item it needs, and asks the server to run it. The server does the work and sends the result back. The model uses the result to write the answer you read.

The important part is who decides. Not you, and not a developer. Nobody writes code for your particular question in advance. The model works out which server to call, and with what, while it is answering.

What a server can offer

The specification defines three kinds of thing a server can offer:

• Tools are actions the assistant can run: look up orders, run a report, update a record.
• Resources are information the assistant can read for context, such as the contents of a file or a database’s layout.
• Prompts are ready-made instructions a server can supply, so a common task is phrased the same way each time.

Most of what you will notice as a user are tools. A server’s list of tools is, in practice, the full list of what it can do for you. If an action is not on the list, the assistant cannot do it through that server.

Local and remote servers

Servers run in one of two places.

A local server runs on your own computer. The app starts it, talks to it directly, and it stops when you close the app. Your login details stay on your machine, and so does the job of installing and updating it.

A remote server runs on the provider’s side and is reached over the internet. There is nothing to install. You add its address in your assistant’s settings, sign in when it asks, and it is there in every conversation afterwards. If you are not a developer, this is usually the easier kind.

What it changes for you

Three things, in practical terms.

Your assistant answers from your own numbers. Instead of a paragraph about how you could find out, you get the figure, taken from the account it lives in, while you are still in the chat.

You are not locked to one assistant. Because every app that supports MCP speaks the same protocol, the docs describe it as a way to “build once and integrate everywhere”. A server you connect in one assistant can be added to another, and it behaves the same way.

The assistant can do work, not only report on it. Tools can change things as well as read them: mark an order as completed, send a message, update a price. Whether a given server allows that, and on what terms, is up to the server.

What MCP does not do

The limits matter as much as the reach.

MCP does not make an answer correct. The model still has to pick the right tool and read the result properly, and a server can only answer from the data it can reach.

MCP does not decide what a server may see or change. The protocol’s documentation says it focuses solely on the exchange itself, not on how an app uses the model or the information. Access is set by the server and by what you grant when you connect an account. That is why the useful question about any server is not “does it use MCP” but “what can it reach, can it change anything, and who approves a change”. We wrote a longer piece on what makes a connector safe to connect.

The words you will see

Different apps use different names for the same setting. Some say MCP server, others say connector. For most purposes they mean the same thing: something you add to your assistant that lets it reach an outside system. The precise difference, and why it matters for questions that span more than one account, is in What are MCP connectors?

Where HeyMetra fits

HeyMetra is a remote MCP server for the accounts you run your business on: your store, your CRM, your subscriptions, your app and search data, and the places you send messages. You connect each account once, in HeyMetra. Then you add one address to the AI assistant you already use, and you can ask it questions that need those accounts, including questions that need two of them at once.

When a question needs a change rather than an answer, HeyMetra does not make it on its own. The change comes back to you first, written out in full, and nothing happens in the account until you approve it.

The assistant pages show where the setting lives in each app, and the connectors page lists every account you can connect.