What Model Context Protocol actually is, why it exists, and how to connect one, in plain English.
Imagine hiring the smartest employee you've ever met. But every time they need to use Gmail, Slack, your CRM, or your database, you have to hand them a completely different instruction manual. And every time you add a new tool, you have to teach them all over again.
That sounds ridiculous, but that's essentially how we've been building AI applications.
MCP is the fix for that. It's why powerful AI integrations are suddenly more accessible than ever.
For decades, APIs have been the backbone of modern software. Check the weather, book a flight, log in with Google, an API is working behind the scenes. One app sends a request, another processes it and sends back an answer. Fast, predictable, reliable.
But for that to work, the first app has to already know three things:
Say you're building a weather app. The endpoint is like a department inside the weather company, it tells the server which thing you want:
api.weathercompany.com/current
But that's not enough, it needs to know which city. That extra info is a parameter:
api.weathercompany.com/current?city=Miami
And the server sends back a response (usually as JSON, a structured way to organize data):
{ "city": "Miami", "temp": "86", "sky": "sunny" }
The catch: the software already knew all the rules before it ever sent the request.
Your weather app already knows the endpoint, the parameters, and the response, because a developer programmed all of it ahead of time. An AI model doesn't have that advantage.
Ask it "what's the weather in Miami today?" and it understands your question, but it doesn't automatically know which endpoint has today's weather, which parameters that endpoint expects, or what the response will look like. All it starts with is your intent.
Given the exact steps to follow.
Starts with a goal, and has to figure out which tool to use and how.
That gap is exactly why MCP was created.
An MCP server sits between an AI model and an application. Its job is to advertise what the application can do, in a standardized format the AI understands.
Think of it as a table of contents for the application.
Instead of forcing the AI to guess what's available, the MCP server tells it: here are the capabilities I offer, here's what each one does, here's the information I need from you, and here's what I'll return. It does that using structured descriptions called JSON schemas.
So instead of a developer manually teaching the AI how to use every endpoint, the application does the job of describing itself.
The APIs didn't change. The app didn't change. What changed is the AI no longer has to be manually taught, the application advertises what it can do.
Most of the time, your existing APIs are still doing the real work behind the scenes: retrieving data, updating databases, sending emails, creating tickets. What changed is who knows how to use them.
Before MCP, developers had to manually teach the AI how to talk to every API. With MCP, the application advertises its own capabilities and the AI discovers and uses them on its own.
It's not MCP versus APIs. It's MCP working with APIs. APIs stay the foundation; MCP just makes them easy for AI to understand and use.
Creating a website once meant writing HTML by hand. Then website builders opened it to millions more people. Then no-code tools. Now it's happening with AI.
MCP doesn't eliminate developers, it dramatically reduces the custom work needed to connect AI to the tools you use every day. That means powerful AI workflows aren't just for big companies with engineering teams, but for solopreneurs, creators, small businesses, anyone with an idea.
The question is shifting from "can I build this?" to "what do I want to build?"
An MCP starts with a small config file in your project, .mcp.json. It simply tells Claude where the MCP server lives (its address) and which project it's for. Claude Code can create this file for you, just ask it to set up the Supabase MCP for your project.
In your terminal, open Claude Code and type /mcp. This is the command that actually connects, it reads your .mcp.json and starts the handshake.
Pick the server (Supabase) and approve it. Most MCPs use a secure sign-in (OAuth), so a browser window opens and you click authorize. That one-time approval is what gives Claude permission to use it.
That's it. Claude can now see the capabilities the MCP advertises and use them on its own. You don't hand it endpoints or parameters, you describe what you want:
My goal isn't to teach you one AI tool after another. It's to help you build an AI-powered web app people actually use and subscribe to, and along the way you'll naturally learn the technologies that make it possible: Claude Code, MCPs, databases, authentication, payments, and whatever comes next.
The tools will change. The destination won't.
Join Leveraging AI โ