Stop 15 · AI Agents & OrchestrationTopics · 8/9

Model Context Protocol (MCP, Overview)

In short

Every AI app used to write its own GitHub plugin, its own Slack plugin, its own database plugin. MCP is one standard so a tool written once works in many AI apps.

An open standard for connecting AI models to external tools and data sources through a consistent protocol — designed to reduce the need for one-off integrations per model or app.

The blob happily plugs one cable into several different gadgets lined up on a desk, next to a tangled heap of old wires.

Think of it like this

One charger for every phone

Remember when every phone had a different charger and the hostel drawer was a tangle of wires? USB-C fixed that: one plug shape, any device. MCP is that plug for AI apps. Build a server once, and any MCP client can plug it in.

Watch it happen

An AI app using an MCP server

You

AI app

MCP server

initialize

1/6On start, the app's MCP client connects to the server and they agree on protocol version and features.

The big ideas

What you need to know

Hosts, clients, servers

The host is the AI app. It runs one MCP client per server. Servers expose capabilities to it.

Tools, resources, prompts

Servers can offer tools to call, resources to read, like files, and prompt templates for users to pick.

stdio or HTTP

Local servers usually talk over stdio as a child process. Remote servers use HTTP, with auth.

Trust servers carefully

A server can run code on your machine. Install only ones you trust and review the tools they expose.

See it in code

Walk through the code

python

1from mcp.server.fastmcp import FastMCP2 3mcp = FastMCP("mess-menu")4 5@mcp.tool()6def todays_menu(meal: str) -> str:7    """Return the hostel mess menu for breakfast, lunch or dinner."""8    menu = {"breakfast": "Poha and chai", "lunch": "Rajma chawal",9            "dinner": "Paneer and roti"}10    return menu.get(meal.lower(), "No menu for that meal")11 12if __name__ == "__main__":13    mcp.run()  # talks over stdio by default

1/3Create a server with a name clients will see.

A complete MCP server with the official Python SDK (pip install mcp). Any MCP client can now call todays_menu.

Save yourself the pain

Common mistakes

Oops

Thinking MCP replaces tool calling.

Do this instead

MCP is a standard way to package and share tools. Under the hood, the model still uses normal tool calling.

Oops

Installing random MCP servers from the internet with full disk access.

Do this instead

Read the code or use trusted sources, and give servers only the folders and keys they need.

Your mission

Run your first MCP server

~30 min

Quick check

You wrote one MCP server for your college timetable. What's the main benefit?

Go deeper

1Read the MCP specification's core concepts: servers, tools, resources.
2Try connecting an MCP-compatible client to a simple MCP server.
3Understand what problem MCP solves compared to custom function-calling integrations.