Build your first MCP server in Python

About me

Python Cloud Advocate at Microsoft

Formerly: UC Berkeley, Coursera, Khan Academy, Google

Find me online at:

About you

Turn to your neighbor and introduce yourself:

• Name
• Location
• Is this your first PyCon, or have you attended before?
• What are you excited for at PyCon 2026?
• Make up a new meaning for "MCP" (e.g. "More Coffee Please") and write it on a sticky note.

AI agents

You've probably used an AI agent...

Coding agents like GitHub Copilot and Claude Code are AI agents.
Chatbots like ChatGPT also have agentic loops now.
You can also build your own agents in Python.

The agentic loop in action

Agents rely on LLM tool-calling

MCP: Model Context Protocol

MCP is the open standard that tells agent runtimes how to discover and connect to the tools they can call.

🔗 MCP specification: modelcontextprotocol.io

MCP architecture

MCP client/server request flow

MCP clients

Coding agent: VS Code + GitHub Copilot

1. Add MCP servers

Install from Extensions:

Or configure .vscode/mcp.json file:

{
  "servers": {
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/"
    }
  }
}

2. Use from Copilot

Once added, tools appear in chat with approvals and structured results.

Coding agent: Claude Code

Claude Code connects terminal-first coding workflows to MCP servers.

claude mcp add \
	--transport http huggingface \
	'https://huggingface.co/mcp?login'

claude mcp list

> what research papers are in 2026
	about vector retrieval?

⏺ huggingface - Paper Search (MCP)
	query: "vector retrieval 2026"
	concise_only: true

120 papers matched the query.
Here are the first 12 results.

## Foundations of Vector Retrieval

Building agents with frameworks

An agent framework coordinates model calls, tool schemas, tool execution, memory or state, and final responses.

Pydantic AI + MCP

from pydantic_ai import Agent
from pydantic_ai.mcp import MCPServerStreamableHTTP
from pydantic_ai.models.openai import OpenAIChatModel

model = OpenAIChatModel("gpt-5.5")
server = MCPServerStreamableHTTP(url="https://learn.microsoft.com/api/mcp")

agent = Agent(model, toolsets=[server],
  system_prompt="Use available tools to answer questions.")

result = await agent.run("How do I use DefaultAzureCredential in Python?")
print(result.output)
					

Exercise time

Exercise 1:
Connect your coding agent to public MCP servers
tinyurl.com/pyconmcp-ex1

🙋🏽‍♀️ 🙋🏻‍♂️ 🙋🏿 🙋🏼‍♀️ 🙋🏾‍♂️ Got a question? Raise your hand, we'll come around!

Python MCP frameworks

FastMCP server skeleton

from fastmcp import FastMCP

mcp = FastMCP("Expenses tracker")

# Define tools, prompts, and resources here...

if __name__ == "__main__":
    mcp.run(transport="streamable-http", host="0.0.0.0", port=8420)
					

🔗 Docs: gofastmcp.com

MCP transports: STDIO vs. HTTP

Learn more: gofastmcp.com/deployment/running-server

Tool with FastMCP

@mcp.tool
async def add_expense(
    date: Annotated[date, "Date of the expense in YYYY-MM-DD format"],
    amount: Annotated[float, "Positive numeric amount of the expense"],
    category: Annotated[Category, "Category label"],
    description: Annotated[str, "Human-readable description of the expense"],
    payment_method: Annotated[PaymentMethod, "Payment method used"],
) -> str:
    """Add a new expense to the expenses.csv file."""
    # Implementation ...

    return f"Added expense: ${amount} for {description} on {date_iso}"
					

🔗 Full example: expenses_tracker.py

Tool signature → JSON schema

@mcp.tool
async def add_expense(
    date: Annotated[date, "YYYY-MM-DD"],
    amount: Annotated[float, "Amount"],
    category: Category,
    description: str,
    payment_method: PaymentMethod,
) -> str:
    """Add a new expense."""
    ...

{
	"name": "add_expense",
	"description": "Add a new expense.",
	"inputSchema": {
		"type": "object",
		"properties": {
			"date": {
				"type": "string",
				"format": "date"
			},
			"amount": {"type": "number"},
			"category": {"type": "string"},
			"description": {"type": "string"},
			"payment_method": {"type": "string"}
		},
		"required": ["date", "amount", "category",
			"description", "payment_method"]
	}
}

Resource with FastMCP

@mcp.resource("resource://expenses")
async def get_expenses_data():
    csv_content = f"Expense data ({len(expenses_data)} entries):\n\n"
    for expense in expenses_data:
        csv_content += (
            f"Date: {expense['date']}, "
            f"Amount: ${expense['amount']}, "
            f"Category: {expense['category']}, "
            f"Description: {expense['description']}, "
            f"Payment: {expense['payment_method']}\n"
        )
    return csv_content
					

🔗 Full example: expenses_tracker.py

Prompt with FastMCP

@mcp.prompt
def analyze_spending_prompt(
    category: str | None = None,
    start_date: str | None = None,
    end_date: str | None = None,
) -> str:
    # Build filters ...
    return f"""Please analyze my spending patterns {filter_text} and provide:
    1. Total spending breakdown by category
    2. Average daily/weekly spending
    3. Most expensive single transaction
    4. Payment method distribution"""
					

🔗 Full example: expenses_tracker.py

Run the server

Start the server:

uv run servers/expenses_tracker.py

Point agent at the local URL:

http://localhost:8000/mcp

Test with MCP Inspector or MCPJam

npx @modelcontextprotocol/inspector

npx @mcpjam/inspector@latest

Useful debugging loop: direct server logs first, Inspector or MCPJam second, LLM client last.

Exercise time

Exercise 3:
Build your own product-store MCP server
tinyurl.com/pyconmcp-ex3

🙋🏽‍♀️ 🙋🏻‍♂️ 🙋🏿 🙋🏼‍♀️ 🙋🏾‍♂️ Got a question? Raise your hand, we'll come around!

MCP as a collaboration protocol

New MCP features turn agents from executors into collaborators.

Elicitations

Elicitations let servers request user input through the client.

Form mode elicitation flow

MCP Apps

MCP Apps let tools return rich, interactive UI that renders inside chat.

MCP Apps flow

Making MCP Apps with FastMCP

Four paths, from Python components to custom UI.

Learn more: gofastmcp.com/apps/overview

Exercise time

Exercise 4:
Add visual UI and purchase confirmation
tinyurl.com/pyconmcp-ex4

🙋🏽‍♀️ 🙋🏻‍♂️ 🙋🏿 🙋🏼‍♀️ 🙋🏾‍♂️ Got a question? Raise your hand, we'll come around!

Three approaches to restricting access

Pick the access model that matches the server boundary and the data risk.

OAuth request/response

MCP client makes requests to MCP server on behalf of a signed-in user:

OAuth 2.1 overview

OAuth 2.1 is a standard for allowing resource owners to make authorized requests. MCP auth builds on top of OAuth 2.1.

OAuth flow for MCP (Simplified)

How the authorization server validates the client

Support for MCP Auth across auth providers

Keycloak: open-source identity server

Runs from a Docker image Admin portal Realms separate tenants Supports OAuth flows and DCR

Implementing MCP auth in FastMCP

Exercise time

Exercise 5:
Add Keycloak authentication to your MCP server
tinyurl.com/pyconmcp-ex5

🙋🏽‍♀️ 🙋🏻‍♂️ 🙋🏿 🙋🏼‍♀️ 🙋🏾‍♂️ Got a question? Raise your hand, we'll come around!

Productionizing MCP servers

• Testing: Test with both direct MCP tools and LLM-driven clients.
• Evaluation: Optimize tool descriptions for user scenarios across agents.
• Usage controls: Rate-limit expensive or state-changing tools.
• Observability: Log tool calls without logging sensitive payloads.
• Security: Continuously audit tools for security risks and logs for abuse.

Security questions to ask every server

• Who is allowed to call this tool?
• What data can the tool read or mutate?
• Does the user understand when state will change?
• Can prompt injection alter the server's authority?
• What happens when the model passes surprising arguments?

Thank you!

Repository:
github.com/pamelafox/pycon2026-mcp-tutorial

Check out more of my MCP talks:

• Improving MCP tool schemas to increase agent reliability
• Identity-aware MCP Servers with Azure Cosmos DB
• An MCP for your Postgres DB
• Python + MCP 3-part series

Before you go