Learn MCP by Building an MCP Server

The hands-on guide to the Model Context Protocol. Build a calculator server, test it the way real AI apps do, upgrade it into a live weather server, and finally build your own AI host — a chat UI where an LLM calls your tools for real.

📖 Full docs: tharunramavath.github.io/learn-mcp-by-building

MCP (Model Context Protocol) is the open standard that lets AI apps (Claude, ChatGPT, VS Code, Cursor) talk to your tools and data — "USB-C for AI". This repo is a from-zero, build-it-yourself tutorial: every concept becomes a file you write, run, and break on purpose. No AI/ML background needed.

Who this is for

• Developers who want to understand MCP by building, not just reading.

• Python users comfortable with functions, async, and JSON.

• People who learn best when they can see the protocol wire traffic fly by.

You do not need any AI/ML background, an API key, or an LLM for Steps 1–6.

Related MCP server: MCP Server Demo

What you'll build

Quickstart (60 seconds)

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
python server_basic.py

That server does nothing visible — by design. It waits for a client. Open a second terminal and talk to it:

python test_client_basic.py

If you see discovery, tools, resources, prompts, and add(2, 3) → 5.0 — you just ran a full MCP conversation. Now go do the steps below.

Table of contents

• Read these docs first

• Project layout

• Step 0 — Setup

• Step 1 — Your first server

• Step 2 — Test it with a tiny client

• Step 3 — Explore with the MCP Inspector

• Step 4 — Attach to a real AI app

• Step 5 — The real-world upgrade

• Step 6 — Under the hood

• Step 7 — Become the host

• Challenges

• FAQ

• Where to go next

• Troubleshooting

Read these docs first

• docs/00-learning-path.md — start here: the recommended reading/doing order

• docs/01-concepts.md — the theory: what MCP is, architecture, primitives

• docs/02-glossary.md — a cheat-sheet of every term

• docs/03-inspector-v2-guide.md — how the new Inspector v2 UI maps to the classic (v1) layout

• docs/04-mcp-vs-alternatives.md — how MCP compares to plain function calling, provider tools, and REST APIs

• docs/05-faq.md — a deeper FAQ than the short one at the bottom of this README

💡 Prefer reading in the browser? The same docs are hosted at tharunramavath.github.io/learn-mcp-by-building.

Project layout

mcp-tutorial/
├── server_basic.py          STEP 1 - your first (offline) server
├── test_client_basic.py     STEP 2 - a tiny client that tests it
├── server_weather.py        STEP 5 - the real-world weather server
├── test_client_weather.py   STEP 5 - a client for the weather server
├── mcp_host.py              STEP 7 - the MCP host/client wrapper (sync bridge)
├── chat_engine.py           STEP 7 - the LLM agent loop (Groq function calling)
├── app.py                   STEP 7 - the Streamlit chat UI
├── run-chat.cmd             STEP 7 - double-click to launch the chat UI
├── claude_desktop_config.json  STEP 4 - config for Claude Desktop (edit the path)
├── .vscode/mcp.json         STEP 4 - config for VS Code
├── inspector.json           STEP 3 - Inspector catalog (launch config for both servers)
├── run-inspector.cmd        STEP 3 - double-click to launch the Inspector
├── requirements.txt
└── docs/
    ├── 00-learning-path.md   START HERE - the sequential reading order
    ├── 01-concepts.md
    ├── 02-glossary.md
    ├── 03-inspector-v2-guide.md
    ├── 04-mcp-vs-alternatives.md
    └── 05-faq.md

Step 0 — Setup (one time)

# From inside the mcp-tutorial/ folder:
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

That installs the official Python MCP SDK (mcp) plus httpx2 (the HTTP client the weather server uses). Verify:

python -c "from mcp.server import MCPServer; print('MCP SDK OK')"

Prefer uv? (faster, modern alternative)

If you have uv installed, you can skip the venv setup entirely — the project has a pyproject.toml:

uv sync            # creates .venv and installs all deps from pyproject.toml
uv run server_basic.py          # run the server
uv run test_client_basic.py     # run the test client

Windows note: use uv for running scripts, but launch the Inspector with npx @modelcontextprotocol/inspector --catalog inspector.json instead of uv run mcp dev — see Step 3. If a GUI app says 'uv' is not recognized, add uv's folder to your PATH permanently: [Environment]::SetEnvironmentVariable("Path", [Environment]::GetEnvironmentVariable("Path","User") + ";$env:USERPROFILE\.local\bin", "User") then fully restart the app.

Step 1 — Your first server: server_basic.py

Open server_basic.py and read it top to bottom. It is ~90 lines and shows all three primitives at once.

from mcp.server import MCPServer

mcp = MCPServer("calculator")        # 1 line to create a server

The 3 tools (what the AI can call)

@mcp.tool()
def add(a: float, b: float) -> float:
    """Add two numbers together.
    Args:
        a: The first number.
        b: The second number.
    """
    return a + b

Three magic things happen here:

1. @mcp.tool() registers the function as an MCP tool.

2. The type hints (a: float, b: float) become the tool's input schema.

3. The docstring becomes the tool's description — this is what the AI model reads to decide when to call it. Write it like the model will read it.

The 1 resource (read-only data)

@mcp.resource("config://app", mime_type="application/json")
def get_app_config() -> dict:
    return APP_CONFIG

Resources have a URI (config://app). The app loads them to give the model context — like opening a file.

The 1 prompt (a template)

@mcp.prompt()
def math_help(problem: str) -> list[dict]:
    return [{"role": "user", "content": f"You are a math tutor. Problem: {problem}"}]

Prompts are templates the user invokes explicitly (like a slash command).

Run it

python server_basic.py

It prints a log line and then just sits there. That is normal — a stdio server waits for a client to spawn it and talk to it.

Why no output? Over the stdio transport, your server's stdout is the protocol. If you print() anything, you corrupt the JSON-RPC messages and the server silently breaks. That's why we use logging (which goes to stderr). This is the golden rule of stdio MCP servers — see docs/01-concepts.md.

Step 2 — Test it with a tiny client: test_client_basic.py

You don't need an AI app yet. This script is a minimal stand-in client that connects to your server over stdio and drives the whole protocol.

python test_client_basic.py

You'll see, in order:

1. Discovery — the client asks what the server supports (server/discover).

2. List tools — notice each tool's auto-generated inputSchema.

3. List resources — the config://app URI.

4. List prompts — math_help.

5. Call tools — add(2, 3) returns 5.0.

6. Read the resource — your config JSON comes back.

7. Get the prompt — the template comes back.

That's the entire MCP conversation, end to end. The client code:

from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

async with stdio_client(server_params) as (read, write):   # spawn server
    async with ClientSession(read, write) as session:      # open conversation
        await session.initialize()                         # handshake
        await session.list_tools()
        await session.call_tool("add", {"a": 2, "b": 3})

Step 3 — Explore with the MCP Inspector (visual UI)

The official MCP Inspector is a web UI for poking at servers and watching the raw JSON-RPC messages.

This project ships an inspector.json catalog that tells the Inspector how to launch both servers using paths relative to this folder (no uv, no PATH lookup, no machine-specific paths).

Easiest way — double-click run-inspector.cmd, or run it from the terminal:

.\run-inspector.cmd

Or invoke npx directly:

npx @modelcontextprotocol/inspector --catalog "<your-path>\mcp-tutorial\inspector.json"

The web UI opens at http://localhost:6274 (the browser auto-launches; the URL may carry an MCP_INSPECTOR_API_TOKEN= query param — keep it if you bookmark the page). Pick a server (calculator or weather) from the catalog and click Connect. Click around:

• Tools / Resources / Prompts tabs — browse what your server exposes.

• Call a tool — try add with {"a": 10, "b": 5}.

• Protocol tab — watch the actual JSON-RPC requests and responses fly by. This is where you see what Step 2 did programmatically.

If the UI looks different from YouTube tutorials, that's because this project runs the new Inspector v2. See docs/03-inspector-v2-guide.md for a feature-by-feature map (transport type, command/args, pings, sampling, roots, protocol routes, etc.) to the v2 tabs.

Windows note: avoid mcp dev / uv run mcp dev — it forces the Inspector to spawn your server via uv, which often fails on Windows (the Inspector's spawned process can't see uv on PATH). The --catalog approach above avoids that entirely by using relative paths to the venv Python. Also launch the Inspector from PowerShell or cmd (double-click run-inspector.cmd) rather than from Git Bash — the MSYS environment mis-resolves node for child processes and the Inspector dies with spawn node ENOENT. Only one Inspector can run at a time — if you see "PORT IS IN USE", close the existing tab/instance instead of starting another. First run downloads the Inspector via npx and takes a minute.

Step 4 — Attach to a real AI app

Now the payoff: plug your server into an AI assistant you already use.

Claude Desktop

Copy claude_desktop_config.json to your real config file: C:\Users\<you>\AppData\Roaming\Claude\claude_desktop_config.json, then fully restart Claude Desktop. Your calculator server will appear. Ask:

"Use the calculator server to multiply 17 by 23."

The config tells Claude Desktop: launch this server by running this Python script, then talk to it over stdio.

VS Code

VS Code reads project-scoped MCP servers from .vscode/mcp.json (already included). Open this folder in VS Code, then:

1. Open the command palette → "MCP: List Servers" / "MCP: Add Server".

2. Or open Copilot Chat and use the MCP tools that appear.

Cursor

Settings → MCP → Add new global MCP server, and paste the same JSON shape from claude_desktop_config.json.

Paths in the config files: inspector.json and .vscode/mcp.json use relative paths (.venv\Scripts\python.exe), so they work as long as you open the project folder. claude_desktop_config.json is a copy-paste example and uses a <YOUR-PATH> placeholder — replace it with the absolute path to your checkout before copying (Claude Desktop doesn't resolve relative paths). On Windows use double backslashes \\ in JSON.

Step 5 — The real-world upgrade: server_weather.py

Now you take what you learned and build something genuinely useful: a weather server for India backed by the free Open-Meteo API (no API key required).

Read server_weather.py. The MCP part is identical to Step 1 — what's new:

• Async tools — async def because we do I/O.

• External API calls — httpx2 fetches live data; make_request wraps it with error handling so a network failure returns a friendly message instead of crashing.

• City name geocoding — users just type Mumbai, not coordinates. geocode_city resolves the name to a latitude/longitude internally via Open-Meteo's geocoding API. A classic real-world API pattern.

• Heuristic alerts — get_alerts derives advisories from current and near-term conditions (thunderstorm, heavy rain, fog, extreme heat), since India has no free keyless official alert feed.

Test it (needs internet):

python test_client_weather.py

Then explore it in the Inspector (the weather server is already in the catalog):

npx @modelcontextprotocol/inspector --catalog "<your-path>\mcp-tutorial\inspector.json"

Then plug it into VS Code / Claude Desktop using the weather entry already in .vscode/mcp.json / claude_desktop_config.json, and ask:

"What's the weather like in Mumbai?"

Step 6 — Under the hood (recommended reading)

Read docs/01-concepts.md fully. The two ideas that make everything click:

1. Two layers. Data layer = the JSON-RPC 2.0 messages (what). Transport layer = how messages travel (stdio locally, Streamable HTTP remotely). Same messages, different plumbing.

2. Stateless + discovery. Every request carries the protocol version and capabilities; the client discovers what the server supports before using it.

Then try the challenges below.

Step 7 — Become the host: chat with your servers (Streamlit UI)

So far your servers just sit and wait for a client. Now you build the other half of MCP — the host (the AI app). This is a real chat UI in your browser that spawns your servers, discovers their tools, and lets an LLM call them through function calling.

Three files, each a distinct layer:

• mcp_host.py — a reusable MCP client/host. It spawns a server over stdio and exposes plain synchronous methods (list_tools, call_tool, read_resource, get_prompt, ...). MCP's Python SDK is async but Streamlit is sync, so it runs a real asyncio event loop on a background thread and bridges the two.

• chat_engine.py — the agent loop. It sends your message plus the discovered tool schemas to Groq. If the model decides it needs a tool, the engine executes it on the server and feeds the result back, repeating until the model answers in plain text. That loop is function calling driving real MCP tools.

• app.py — the Streamlit UI. Pick a server, paste a Groq API key, connect, and chat. Every reply can show an agent trace of the real MCP tool calls that produced it.

Run it — double-click run-chat.cmd, or:

python -m streamlit run app.py

The UI opens at http://localhost:8501. You need a free Groq API key (https://console.groq.com/keys). Then:

1. Paste the key in the sidebar.

2. Pick the calculator server → Connect. The sidebar lists its discovered 3 tools, 1 resource, and 1 prompt.

3. Ask "What is 17 * 23?". The agent trace under the reply shows the LLM called your real multiply tool and used its answer.

4. Switch to the weather server and ask "What's the weather in Mumbai?" — it geocodes the city and fetches live data, all via MCP.

Why this is the big finale: every AI app is a host. Claude Desktop, VS Code, and Cursor do exactly what app.py does — spawn servers, discover tools, and call them on the model's behalf. You just built one from scratch.

Challenges (do these yourself)

1. Add a divide tool to server_basic.py. What happens when b == 0?

2. Add a resource template — a resource with a URI like weather://city/{city} instead of a fixed URI.

3. Make server_basic.py a remote server — change the last line to mcp.run(transport="streamable-http") and see what URL it prints. (You'll need to configure a client for HTTP transport to talk to it.)

4. Write your own tiny client that takes a math question in English, extracts numbers, and calls your calculator tools.

5. Add a tool to server_basic.py and re-connect the chat UI — the host should discover it automatically. What happens if the LLM divides by zero?

6. Swap the Groq model in the chat sidebar. Which models call tools more reliably? (Watch the agent trace to see.)

FAQ

Do I need an API key or an AI model to follow along? No. Steps 1–6 need only Python. The only key in the whole project is the free Groq key used in Step 7 (the host). You can go through the entire protocol without ever touching an LLM.

Why does my server "do nothing" when I run it? It's working. A stdio server doesn't run on its own — it waits for a client to spawn it and speak JSON-RPC over stdin/stdout. See Step 2 / Step 3.

What's the difference between a tool, a resource, and a prompt? Short version: tools do something (the model calls them), resources are data to read (the app loads them), prompts are recipes (the user invokes them). Full details in docs/01-concepts.md.

Isn't this just "function calling"? Function calling is how an LLM decides to use a tool — MCP is the standard for transporting tools and data between apps and models. They're complementary, not the same thing. The whole comparison is in docs/04-mcp-vs-alternatives.md.

stdio vs Streamable HTTP — which do I use? Same JSON-RPC messages, different plumbing. stdio for local servers you run yourself; Streamable HTTP for remote/deployed servers. See docs/01-concepts.md.

Can I plug my server into ChatGPT? Claude Desktop, VS Code, and Cursor are the easy wins in Step 4. Support varies by app — any app that speaks MCP can use your server; that's the whole point.

Do I have to learn JSON-RPC? No. The SDK generates and parses it for you — but Step 3 shows you the raw traffic so you understand what's happening underneath.

More questions answered in docs/05-faq.md.

Where to go next

• Official docs: https://modelcontextprotocol.io

• MCP specification: https://modelcontextprotocol.io/specification/latest

• Reference servers: https://github.com/modelcontextprotocol/servers

Troubleshooting