7 min read

What is MCP? And how does it differ from an API?

Blue Hello Growth cover with the text MCP vs API, below it Claude, ChatGPT and Claude Code

In short

  • With an MCP connection you link Claude, Claude Code or ChatGPT to another service, such as your mail, Notion or Search Console. You add it yourself, without programming.
  • An API is the connection a system offers to software. An MCP is a standard on top of such a system, made for AI tools. Underneath an MCP there is usually an ordinary API.
  • For your daily work the difference often matters little. What matters more is that you have connections, and that you know which permissions they get.

These days I use all kinds of connections to third-party services in Claude. And I ran into the same question you probably have now: do I connect an API here, or is there an MCP? Sometimes it is an API, sometimes an MCP. This is what I wish I had known when I started.

The short answer: an MCP is the standard you use to attach a service to ChatGPT or Claude. An API is that service's connection for software in general. You usually have both, because under the MCP sits an API. For you as a user, what matters most is that you can make that connection yourself, in Claude and ChatGPT with a few clicks.

What is an MCP connection in ChatGPT or Claude?

MCP stands for Model Context Protocol. According to the official documentation, it is an open-source standard for connecting AI applications such as Claude or ChatGPT to external systems: data sources such as files and databases, tools such as search engines, and workflows such as specialised prompts. The documentation compares it to a USB-C port: one standard way to plug something in.

In practice you get a connection. Claude calls it a connector, ChatGPT a plugin. The service you want to work with runs an MCP server: the part on its side that follows the standard. You connect your AI tool to it. After that Claude or ChatGPT can answer questions and take actions in that service.

How to make the connection

In Claude

In Claude an MCP connection is called a connector. You find them under Customize and then Connectors. Choose Discover, search for the service, select Connect to Claude, sign in to that service and approve the access. If the service is not in the list and you have the server's URL, you can add it yourself. On Team and Enterprise plans an Owner adds a connector for the organisation, after which members connect with their own account. See Get started with connectors.

In Claude Code

Claude Code is Claude in your terminal, used mostly by people who work with code. You add an MCP server there with one command, for example for a remote server:

claude mcp add --transport http notion https://mcp.notion.com/mcp

You then sign in through /mcp. You also choose whether the connection applies only to you, to a project or to all your projects. See Connect Claude Code to tools via MCP.

In ChatGPT

In ChatGPT you add an MCP server through plugins: choose the plus button, then Add custom MCP server, enter the URL and the authentication details, read the risk warning and create it as a plugin. You then use it in chat and in deep research. OpenAI's documentation names no plan, so check in your own ChatGPT whether you see this option. See Building MCP servers for plugins and API integrations.

Copilot uses the same standard for its connections, as you can read in our article on Copilot Cowork.

My most used MCP connection: Outlook

The MCP connection I use most myself is the one with Outlook. Claude can look into my inbox, so I have it read through mail quickly, update projects based on what comes in and run analyses on my inbox. That is a lot faster than searching and reading it all myself.

It is also the example where the risks further down matter most. An inbox is full of other people's text, and Claude works with my permissions. More on that under "What do you arrange before connecting anything?".

MCP vs API: what is the difference?

An API (application programming interface) is a system's front door for other software. Picture a counter with forms: you need to know exactly which form to fill in and where to hand it in. A developer reads the documentation and writes code that does it. That works well for something that does the same thing every time, such as an order that goes to the warehouse. But someone has to build it.

An MCP server hands the AI a menu: these are the things I can do, and this is what each one does. Technically that is called tools, resources and prompts. The AI reads the menu and chooses what fits your question. Counter and menu are our own pictures, not terms from the documentation.

APIMCP server
What is it?A system's connection for softwareA layer on top of that system, for AI tools
Who is it for?DevelopersUsers of Claude, ChatGPT or Claude Code
Who makes the connection?A developer, with codeYou, in your AI tool
Who decides the call?The code that was writtenThe AI, per question
How does the AI know what is possible?It is written in the codeThe server lists its actions with descriptions
FitsFixed processesChanging questions in a conversation
ExampleNightly export to accounting"Which pages lost the most visitors?"

Say you attach Search Console to Claude and ask which pages lost the most search traffic last month. Claude picks the right actions and retrieves the data, and a follow-up question can come straight after. With an API alone, someone first has to build a script or export for that one question. This is an example of the principle, not a measurement.

Does the difference actually matter?

For many people it does not. If you just want Claude or ChatGPT to reach your mail, calendar or documents, you look for a connection and it does not matter whether an API or an MCP sits underneath. If the service has a connector or MCP server, use it. If it only has an API, someone usually has to build a connection for it.

The difference only starts to count in two situations:

  • A fixed process that does the same thing every time, such as an invoice that goes to accounting. The AI does not need to choose. An API or a fixed automation is more predictable and easier to check.
  • An AI agent that takes steps on its own. An AI agent uses tools to do its work. MCP is one way to give it those tools, and under those tools sit APIs.

What do you arrange before connecting anything?

Permissions. A connection works with your permissions. If the AI may write in your CRM, it can also change records there. Start read-only and switch on writing only for a task that has proved itself.

Content as instructions. An AI that reads text from mail or documents can come across commands that did not come from you. Keep a person in the loop for actions that change, send or delete something. Also read the MCP security page.

Who runs the server? A remote server sees the data you give it. Check who runs it and where that data goes, especially with client data.

Now you get it

An API is the door of a system, meant for code. An MCP server is the menu on top of that door, meant for an AI. When you hear that someone has to connect an API, something is usually being built whose behaviour is fixed in advance. When you hear MCP, you attach a service to Claude or ChatGPT yourself and then talk to it.

Usually there is an API underneath and an MCP on top. Start with one service you look things up in often, connect it read-only and see what it gives you.

Sources

Newsletter

The AI news that matters, once a month

One email a month with the most important developments in AI and what they mean for your organisation. No noise.

You will receive the Hello Growth newsletter, about once a month. First you get an email to confirm your subscription. You can unsubscribe at any time with the link under every newsletter. Read how we handle your data.

Let us begin

Want this in your own organisation?

We build it, or we teach your team to. Either works.