If you’ve been following AI news, you’ve probably seen “MCP” everywhere. And if you’ve ever worked with software, you already know “API.” They sound like they do the same thing: let software talk to other software. So what’s actually different, and why does it matter for AI?
First, the quick definitions
- API (Application Programming Interface): a set of doors into a service. Each door (an “endpoint”) does one thing, like “create a contact” or “send an invoice.” APIs have been around for decades and power almost every app you use.
- MCP (Model Context Protocol): a newer open standard, introduced in late 2024, built specifically so AI agents can connect to tools. An MCP server sits in front of a service and tells the AI, in plain language, what it can do and how to do it.
The handshake analogy
Imagine you walk up to someone at a networking event and shake their hand. You’ve connected, and you both agree to engage. Then they say nothing. You don’t know their name, what they do or what language they speak. You can’t really talk to them until you figure that out on your own.
That’s an API. The connection works, but the AI has no built-in idea what the service can do. Someone has to hand it that knowledge first: which endpoints exist, what each one does and what information it needs.
Now picture the same handshake, except this time the person says: “Hi, I’m Carlos. I run a web agency, I speak English and Spanish, and here’s what I can help you with.” Now you know exactly how to talk to them.
That’s MCP. The moment an AI connects, the server introduces itself: here are my tools, here’s what each one does, here’s what I need from you.
Side by side
| API | MCP | |
|---|---|---|
| Built for | Developers and apps | AI agents |
| How the AI learns what it can do | Someone provides docs or writes code | The server describes itself on connect |
| Setup per tool | Custom work for each integration | Connect once, works across AI apps |
| When the service changes | Integration may need updating | New tools show up automatically |
| How common | Nearly every service has one | Growing fast, but not everywhere yet |
Here’s the part people miss: they work together
MCP doesn’t replace APIs. Most MCP servers are built on top of an existing API. MCP is the friendly introduction layered over it.
That also means no MCP doesn’t mean no AI. Plenty of tools don’t have an MCP server yet. AI agents like Viktor, and others like it, can still connect through the API directly. They read the documentation, write the code to make the calls and get the job done. In handshake terms, someone whispered the other person’s name and background in the agent’s ear before it walked over.
The bottom line
An API gives your AI a handshake. MCP gives it a handshake and an introduction. MCP is the easiest path when it’s available. When it isn’t, a well-documented API and a capable agent still get you there.
