MCP Explained Simply: What It Is, Why It Matters, and How to Build Your First Server

Jay Kumar — Fri, 17 Jul 2026 03:10:37 +0000

I have worked in cloud and platform engineering for 16 years. When I first heard "agentic AI" and "MCP", it sounded complex. It is not. This post explains everything in simple words. You do not need to read anything else to understand MCP. There is a glossary at the bottom if any word is new to you.

The problem MCP solves

AI tools like Claude and ChatGPT are smart, but they are locked in a box. They cannot see your files. They cannot read your database. They cannot use your company tools.

Before MCP, if you wanted to connect an AI app to a system, you had to write custom code for each pair. Ten AI apps and ten systems means one hundred custom integrations. This does not scale. I have seen this exact problem block AI projects in large companies.

What is MCP?

MCP stands for Model Context Protocol. It is an open standard. Think of it like a USB-C port for AI.

Before USB-C, every device had its own cable. Now one cable works for everything. MCP does the same thing for AI. You build one MCP server for your system, and every AI app can use it. You build one AI app with MCP support, and it can connect to every system.

MCP was created by Anthropic and is now an open project under the Agentic AI Foundation (AAIF), which is part of the Linux Foundation. That means no single company owns it. Anyone can use it for free.

The architecture

Read the picture from top to bottom:

Layer 1 — AI applications. These are the apps people use: Claude, ChatGPT, VS Code, Cursor, or your own company chatbot. In MCP language these are called "hosts".

Layer 2 — The MCP protocol. This is the shared language. Every connection uses the same rules. An MCP server can offer three things: tools (actions the AI can do), resources (data the AI can read), and prompts (ready-made instructions).

Layer 3 — MCP servers. These are small programs. Each one connects to one system: your files, a database, Slack, GitHub, or your own custom tool. You write it once and every AI app can use it.

Layer 4 — Your systems. Your data stays where it is. MCP does not move your data. It only gives the AI a safe, controlled way to reach it.

The key idea: without MCP, every AI app times every system means custom work. With MCP, you build each side once and connect everything.

Build your first MCP server in 20 minutes

Now the fun part. We will build a small MCP server in Python. If you can write a simple Python function, you can do this.

Our server is a "Community Toolkit". It gives the AI three simple tools: count days until an event, write a meetup announcement, and write a thank-you note.

Step 1 — Install the SDK (1 minute)

pip install "mcp[cli]"

Step 2 — Write the server (10 minutes)

Create a file called server.py:

from datetime import date
from mcp.server.fastmcp import FastMCP

# Create the server. The name shows up in the AI app.
mcp = FastMCP("community-toolkit")

@mcp.tool()
def days_until_event(event_date: str) -> str:
    """Count days until a community event. Date format: YYYY-MM-DD."""
    target = date.fromisoformat(event_date)
    delta = (target - date.today()).days
    if delta < 0:
        return f"That event was {-delta} days ago."
    if delta == 0:
        return "The event is today!"
    return f"{delta} days until the event."

@mcp.tool()
def meetup_announcement(event_name: str, event_date: str,
                        location: str, signup_url: str) -> str:
    """Generate a ready-to-post meetup announcement in Markdown."""
    return (
        f"# {event_name}\n\n"
        f"**When:** {event_date}\n"
        f"**Where:** {location}\n\n"
        f"Join fellow builders for talks, demos, and community time.\n\n"
        f"[Register here]({signup_url})\n"
    )

if __name__ == "__main__":
    mcp.run()

That is a complete MCP server. Look how simple it is. The @mcp.tool() line turns a normal Python function into a tool the AI can use. The docstring (the text in triple quotes) is very important. The AI reads it to decide when to use your tool.

Step 3 — Test it (5 minutes)

mcp dev server.py

This opens a test page in your browser called the MCP Inspector. You can call your tools by hand and see the results. No AI needed yet. This is the fastest way to check your work.

Step 4 — Connect it to Claude Desktop (5 minutes)

Open the file claude_desktop_config.json (in Claude Desktop: Settings → Developer → Edit Config) and add:

{
  "mcpServers": {
    "community-toolkit": {
      "command": "python",
      "args": ["/full/path/to/server.py"]
    }
  }
}

Restart Claude Desktop. Then ask: "How many days until my meetup on 2026-09-15?" Claude will find your tool, call it, and give you the answer. Watching this work the first time feels like magic.

Full code is here: github.com/jaykuma007/mcp-first-server

Three mistakes to avoid

I made all of these, so you do not have to:

Use the full file path. In the Claude Desktop config, a short path like server.py fails silently. Use the complete path like /Users/you/code/server.py.
Restart the app. Claude Desktop reads the config only when it starts. If your server does not show up, restart the app.
Write clear docstrings. The AI picks tools based on your descriptions, not your code. A vague docstring means your tool never gets used.

A note for enterprise readers

If you work in a large company, here is why MCP matters to you. Your AI strategy will need many connections: data warehouses, internal APIs, identity systems, monitoring. Without a standard, each connection is a one-off project and a security review nightmare. With MCP, you get one pattern to build, one pattern to secure, and one pattern to audit. You can put your MCP servers behind your normal access controls, and your data never leaves your systems.

What is AAIF?

The Agentic AI Foundation (AAIF) is the open home for agentic AI projects, under the Linux Foundation. It hosts:

MCP — the protocol you just learned
Goose — an open-source AI agent for developers
AGENTS.md — a simple standard file that tells AI coding agents how to work with your project
agentgateway — a gateway for managing agent traffic

AAIF also runs an Ambassador Program. Ambassadors create one public contribution per month — a tutorial, talk, video, or blog post — that helps developers use these projects. This post is my first contribution on that path. Applications are open all year at aaif.io/ambassadors, and cohorts are selected in June and January.

What is next

This is post one of a series. Next I will connect an MCP server to a real API, and later I will cover how to run MCP servers in production: deployment, security, and monitoring — the platform engineering side that most tutorials skip.

If this post helped you, star the repo, and try building one small tool for your own workflow. One afternoon is enough.

Glossary

Agent / Agentic AI — An AI that does not just answer questions. It can take actions: call tools, read data, and complete multi-step tasks.

AAIF (Agentic AI Foundation) — The open, neutral home for agentic AI projects like MCP, Goose, and AGENTS.md. Part of the Linux Foundation.

Anthropic — The AI company that created Claude and originally created MCP.

Claude / Claude Desktop — An AI assistant by Anthropic. The desktop app can connect to MCP servers.

Docstring — The description text inside a Python function, written in triple quotes. In MCP, the AI reads it to decide when to use the tool.

Host — The AI application the user talks to (Claude Desktop, ChatGPT, VS Code). The host contains an MCP client.

JSON-RPC — The simple message format MCP uses under the hood. You do not need to touch it; the SDK handles it.

Linux Foundation — A nonprofit that hosts major open-source projects (Linux, Kubernetes, and now AAIF).

MCP (Model Context Protocol) — The open standard that connects AI applications to external systems. The "USB-C for AI".

MCP Client — The part inside the host app that talks to MCP servers.

MCP Inspector — A browser tool (mcp dev server.py) for testing your MCP server without an AI.

MCP Server — A small program that exposes one system (files, database, API) to AI apps through MCP.

Open standard — A public specification anyone can use and implement for free. No single company controls it.

Prompt (in MCP) — A reusable instruction template an MCP server can offer to the AI app.

Resource (in MCP) — Data an MCP server lets the AI read, like a file or a database record.

SDK (Software Development Kit) — A code library that does the hard work for you. The MCP Python SDK is installed with pip install "mcp[cli]".

Tool (in MCP) — An action the AI can perform through your MCP server, like "send a message" or "count days".

I am a platform architect in Tokyo, writing simple MCP and agentic AI tutorials for developers in Japan and APAC. Follow me for the next post in this series.

DEV Community: Jay Kumar