🧠 What is MCP Server? (My Simple Understanding)

3 min readMay 17, 2025

“I’m not an expert in AI, but recently I came across something cool while watching a video on the Chai aur Code YouTube channel — https://youtu.be/dZyQNy3-HjU?si=zkQ6BspDUpcHk-qB.
So thought I’d try to explain it in my own words.”

🛠️ What’s the Problem?

If you’ve ever used ChatGPT or any AI coding tool, you probably gave it a prompt like:

“Hey, generate a Python API with login and register routes.”

It gives you code — great!
But then what?

You still have to copy-paste it into your project
Install stuff manually
Fix any broken imports
And AI has no idea what’s already in your code

This is where I got curious…
What if AI tools could actually read your codebase, understand what’s there, and then make changes directly like a teammate?

💡 That’s where MCP comes in.

MCP = Model Context Protocol

It’s not software, not an app you install.
It’s a standard — basically a rulebook for how AI tools (like Cursor or Copilot Workspace) can talk to your files, run commands, and understand your whole setup.

💬 Let Me Explain It My Way:

Imagine you’re working on a Node.js project.
You open Cursor (which looks like VS Code), and say:

“Create an Express server with one /health endpoint."

Without MCP (Old Style):

The AI gives you some sample code
You copy it
Paste it manually
Create the file yourself
Maybe forget to install express

Basically: AI suggests, you do the work.

With MCP:

The AI sees your current project structure
It checks if express is installed
Creates the file at the right place
Adds the route
Even runs npm install express for you

🤯 Feels like a real junior developer working with you.

📦 So what exactly is an MCP server?

Think of it like the “middleman” between your AI model (like Claude or GPT-4) and your local environment.

It knows how to:

Read files in your repo
Edit and write code
Run shell commands
Check logs or errors
Even connect to tools like Slack, GitHub, Figma, etc.

👉 Cursor, Continue.dev, Copilot Workspace — these tools use something like an MCP server behind the scenes.

🧰 What kind of things can MCP servers handle?

MCP uses “tools” like:

File Tools -> Create, delete, rename, edit files
Shell Tools -> run terminal commands
Docs Tools -> search documentation
Web Tools -> make API calls

Think of these like small plugins.
One tool just creates files. Another reads logs. Another checks errors.
Like microservices — but for AI.

💡 Why is this needed?

Right now, when AI interacts with your environment, there’s no common format or rulebook.
Everyone builds their own custom system.

MCP solves this by saying:
“Let’s all use the same protocol, so AI tools work with any editor or backend.”

Kinda like how REST API standardized web services — MCP does the same for AI + developer tools.

🔐 Bonus Example: Security Context

Let’s say your LLM is trying to fix an error in your repo.

You don’t want it accessing your company’s Slack or GitHub issues unless you allow it.

MCP defines context boundaries — like:

What data the AI is allowed to see
What actions it can take (read-only? write?)
Which tools are available
How to “talk” to those tools

So it’s safe, limited, and predictable — not some random script with full access.

🧠 MCP Has 5 Main Context Types

Just for nerds like me who want to go a bit deeper:

Tools — Functions like createFile, readLogs
Resources — Actual files (e.g., logs, CSVs)
Sampling — When AI needs to talk to another model
Prompts — Templates you give to the model
Protocols — Rules about how AI communicates

📌 Summary (in simple words)

MCP is a rulebook, not a tool
It helps AI tools act like a real assistant, not just a chatbot
You’ll see it behind tools like Cursor, Copilot Workspace, Continue.dev
It’s made by Anthropic, and still evolving

🎥 Credit Where It’s Due

I learned about all this from a great video by Chai aur Code — Watch here.
I’m still exploring this space and just wanted to share what I understood.
Let me know if I explained it well — or if something is missing!