# Introduction
The Model Context Protocol (MCP) has changed how large language models (LLMs) interact with external tools, data sources, and services. However, building MCP servers from scratch traditionally required navigating complex boilerplate code and detailed protocol specifications. FastMCP eliminates this roadblock, providing a decorator-based, Pythonic framework that enables developers to build production-ready MCP servers and clients with minimal code.
In this tutorial, you’ll learn how to build MCP servers and clients using FastMCP, which is comprehensive and complete with error handling, making it ideal for both beginners and intermediate developers.
// Prerequisites
Before starting this tutorial, make sure you have:
- Python 3.10 or higher (3.11+ recommended for better async performance)
- pip or uv (uv is recommended for FastMCP deployment and is required for the CLI tools)
- A code editor (I’m using VS Code, but you can use any editor of your choice)
- Terminal/Command Line familiarity for running Python scripts
It is also beneficial to have good Python programming knowledge (functions, decorators, type hints), some understanding of async/await syntax (optional, but helpful for advanced examples), familiarity with JSON and REST API concepts, and basic command-line terminal usage.
Before FastMCP, building MCP servers required you to have a deep understanding of the MCP JSON-RPC specification, extensive boilerplate code for protocol handling, manual connection and transport management, and complex error handling and validation logic.
FastMCP addresses these issues with intuitive decorators and a simple, Pythonic API, enabling you to focus on business logic rather than protocol implementation.
# What is the Model Context Protocol?
The Model Context Protocol (MCP) is an open standard created by Anthropic. It provides a universal interface for AI applications to securely connect with external tools, data sources, and services. MCP standardizes how LLMs interact with external systems, much like how web APIs standardized web service communication.
// Key Characteristics of MCP
- Standardized Communication: Uses JSON-RPC 2.0 for reliable, structured messaging
- Bidirectional: Supports both requests from clients to servers and responses back
- Security: Built-in support for authentication and authorization patterns
- Flexible Transport: Works with any transport mechanism (stdio, HTTP, WebSocket, SSE)
// MCP Architecture: Servers and Clients
MCP follows a clear client-server architecture:
Image by Author
- MCP Server: Exposes capabilities (tools, resources, prompts) that external applications can use. Think of it as a backend API specifically designed for LLM integration.
- MCP Client: Embedded in AI applications (like Claude Desktop, Cursor IDE, or custom applications) that connect to MCP servers to access their resources.
// Core Components of MCP
MCP servers expose three primary types of capabilities:
- Tools: Executable functions that LLMs can call to perform actions. Tools can query databases, call APIs, perform calculations, or trigger workflows.
- Resources: Read-only data that MCP clients can fetch and use as context. Resources might be file contents, configuration data, or dynamically generated content.
- Prompts: Reusable message templates that guide LLM behavior. Prompts provide consistent instructions for multi-step operations or specialized reasoning.
# What is FastMCP?
FastMCP is a high-level Python framework that simplifies the process of building both MCP servers and clients. Created to reduce development headaches, FastMCP possesses the following characteristics:
- Decorator-Based API: Python decorators (
@mcp.tool,@mcp.resource,@mcp.prompt) eliminate boilerplate - Type Safety: Full type hints and validation using Python’s type system
- Async/Await Support: Modern async Python for high-performance operations
- Multiple Transports: Support for stdio, HTTP, WebSocket, and SSE
- Built-in Testing: Easy client-server testing without subprocess complexity
- Production Ready: Features like error handling, logging, and configuration for production deployments
// FastMCP Philosophy
FastMCP relies on three core principles:
- High-level abstractions: Less code and faster development cycles
- Simple: Minimal boilerplate allows focus on functionality over protocol details
- Pythonic: Natural Python idioms make it familiar to Python developers
# Installation
Start by installing FastMCP and the necessary dependencies. I recommend using uv.
If you don’t have uv, install it with pip:
Or install FastMCP directly with pip:
Verify that FastMCP is installed:
python -c "from fastmcp import FastMCP; print('FastMCP installed successfully')"
# Building Your First MCP Server
We will create a practical MCP server that demonstrates tools, resources, and prompts. We’ll build a Calculator Server that provides mathematical operations, configuration resources, and instruction prompts.
// Step 1: Setting Up the Project Structure
We first need to create a project directory and initialize your environment. Create a folder for your project:
Then navigate into your project folder:
Initialize your project with the necessary files:
// Step 2: Creating the MCP Server
Our Calculator MCP Server is a simple MCP server demonstrating tools, resources, and prompts. Inside your project folder, create a file named calculator_server.py and add the following code.
import logging
import sys
from typing import Dict
from fastmcp import FastMCP
# Configure logging to stderr (critical for MCP protocol integrity)
logging.basicConfig(
level=logging.DEBUG,
format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
stream=sys.stderr
)
logger = logging.getLogger(__name__)
# Create the FastMCP server instance
mcp = FastMCP(name="CalculatorServer")
The server imports FastMCP and configures logging to stderr. The MCP protocol requires that all output, except protocol messages, be directed to stderr to avoid corrupting communication. The FastMCP(name="CalculatorServer") call creates the server instance. This handles all protocol management automatically.
Now, let’s create our tools.
@mcp.tool
def add(a: float, b: float) -> float:
"""
Add two numbers together.
Args:
a: First number
b: Second number
Returns:
Sum of a and b
"""
try:
result = a + b
logger.info(f"Addition performed: {a} + {b} = {result}")
return result
except TypeError as e:
logger.error(f"Type error in add: {e}")
raise ValueError(f"Invalid input types: {e}")
@mcp.tool
def subtract(a: float, b: float) -> float:
"""
Subtract b from a.
Args:
a: First number (minuend)
b: Second number (subtrahend)
Returns:
Difference of a and b
"""
try:
result = a - b
logger.info(f"Subtraction performed: {a} - {b} = {result}")
return result
except TypeError as e:
logger.error(f"Type error in subtract: {e}")
raise ValueError(f"Invalid input types: {e}")
We have defined functions for addition and subtraction. Both are wrapped in a try-catch block to raise value errors, log the information, and return the result.
@mcp.tool
def multiply(a: float, b: float) -> float:
"""
Multiply two numbers.
Args:
a: First number
b: Second number
Returns:
Product of a and b
"""
try:
result = a * b
logger.info(f"Multiplication performed: {a} * {b} = {result}")
return result
except TypeError as e:
logger.error(f"Type error in multiply: {e}")
raise ValueError(f"Invalid input types: {e}")
@mcp.tool
def divide(a: float, b: float) -> float:
"""
Divide a by b.
Args:
a: Dividend (numerator)
b: Divisor (denominator)
Returns:
Quotient of a divided by b
Raises:
ValueError: If attempting to divide by zero
"""
try:
if b == 0:
logger.warning(f"Division by zero attempted: {a} / {b}")
raise ValueError("Cannot divide by zero")
result = a / b
logger.info(f"Division performed: {a} / {b} = {result}")
return result
except (TypeError, ZeroDivisionError) as e:
logger.error(f"Error in divide: {e}")
raise ValueError(f"Division error: {e}")
Four decorated functions (@mcp.tool) expose mathematical operations. Each tool includes:
- Type hints for parameters and return values
- Comprehensive docstrings (MCP uses these as tool descriptions)
- Error handling with try-except blocks
- Logging for debugging and monitoring
- Input validation
Let’s move on to building resources.
@mcp.resource("config://calculator/settings")
def get_settings() -> Dict:
"""
Provides calculator configuration and available operations.
Returns:
Dictionary containing calculator settings and metadata
"""
logger.debug("Fetching calculator settings")
return {
"version": "1.0.0",
"operations": ["add", "subtract", "multiply", "divide"],
"precision": "IEEE 754 double precision",
"max_value": 1.7976931348623157e+308,
"min_value": -1.7976931348623157e+308,
"supports_negative": True,
"supports_decimals": True
}
@mcp.resource("docs://calculator/guide")
def get_guide() -> str:
"""
Provides a user guide for the calculator server.
Returns:
String containing usage guide and examples
"""
logger.debug("Retrieving calculator guide")
guide = """
1. **add(a, b)**: Returns a + b
Example: add(5, 3) = 8
2. **subtract(a, b)**: Returns a - b
Example: subtract(10, 4) = 6
3. **multiply(a, b)**: Returns a * b
Example: multiply(7, 6) = 42
4. **divide(a, b)**: Returns a / b
Example: divide(20, 4) = 5.0
## Error Handling
- Division by zero will raise a ValueError
- Non-numeric inputs will raise a ValueError
- All inputs should be valid numbers (int or float)
## Precision
The calculator uses IEEE 754 double precision floating-point arithmetic.
Results may contain minor rounding errors for some operations.
"""
return guide
Two decorated functions (@mcp.resource) provide static and dynamic data:
config://calculator/settings: Returns metadata about the calculatordocs://calculator/guide: Returns a formatted user guide- URI format distinguishes resource types (convention:
type://category/resource)
Let’s build our prompts.
@mcp.prompt
def calculate_expression(expression: str) -> str:
"""
Provides instructions for evaluating a mathematical expression.
Args:
expression: A mathematical expression to evaluate
Returns:
Formatted prompt instructing the LLM how to evaluate the expression
"""
logger.debug(f"Generating calculation prompt for: {expression}")
prompt = f"""
Please evaluate the following mathematical expression step by step:
Expression: {expression}
Instructions:
1. Break down the expression into individual operations
2. Use the appropriate calculator tool for each operation
3. Follow order of operations (parentheses, multiplication/division, addition/subtraction)
4. Show all intermediate steps
5. Provide the final result
Available tools: add, subtract, multiply, divide
"""
return prompt.strip()
Finally, add the server startup script.
if __name__ == "__main__":
logger.info("Starting Calculator MCP Server...")
try:
# Run the server with stdio transport (default for Claude Desktop)
mcp.run(transport="stdio")
except KeyboardInterrupt:
logger.info("Server interrupted by user")
sys.exit(0)
except Exception as e:
logger.error(f"Fatal error: {e}", exc_info=True)
sys.exit(1)
The @mcp.prompt decorator creates instruction templates that guide LLM behavior for complex tasks.
Error handling best practices included here are:
- Specific exception catching (TypeError, ZeroDivisionError)
- Meaningful error messages for users
- Detailed logging for debugging
- Graceful error propagation
// Step 3: Building the MCP Client
In this step, we will demonstrate how to interact with the Calculator MCP Server that we created above. Create a new file named calculator_client.py.
import asyncio
import logging
import sys
from typing import Any
from fastmcp import Client, FastMCP
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
stream=sys.stderr
)
logger = logging.getLogger(__name__)
async def main():
"""
Main client function demonstrating server interaction.
"""
from calculator_server import mcp as server
logger.info("Initializing Calculator Client...")
try:
async with Client(server) as client:
logger.info("✓ Connected to Calculator Server")
# DISCOVER CAPABILITIEs
print("\n" + "="*60)
print("1. DISCOVERING SERVER CAPABILITIES")
print("="*60)
# List available tools
tools = await client.list_tools()
print(f"\nAvailable Tools ({len(tools)}):")
for tool in tools:
print(f" • {tool.name}: {tool.description}")
# List available resources
resources = await client.list_resources()
print(f"\nAvailable Resources ({len(resources)}):")
for resource in resources:
print(f" • {resource.uri}: {resource.name or resource.uri}")
# List available prompts
prompts = await client.list_prompts()
print(f"\nAvailable Prompts ({len(prompts)}):")
for prompt in prompts:
print(f" • {prompt.name}: {prompt.description}")
# CALL TOOLS
print("\n" + "="*60)
print("2. CALLING TOOLS")
print("="*60)
# Simple addition
print("\nTest 1: Adding 15 + 27")
result = await client.call_tool("add", {"a": 15, "b": 27})
result_value = extract_tool_result(result)
print(f" Result: 15 + 27 = {result_value}")
# Division with error handling
print("\nTest 2: Dividing 100 / 5")
result = await client.call_tool("divide", {"a": 100, "b": 5})
result_value = extract_tool_result(result)
print(f" Result: 100 / 5 = {result_value}")
# Error case: division by zero
print("\nTest 3: Division by Zero (Error Handling)")
try:
result = await client.call_tool("divide", {"a": 10, "b": 0})
print(f" Unexpected success: {result}")
except Exception as e:
print(f" ✓ Error caught correctly: {str(e)}")
# READ RESOURCES
print("\n" + "="*60)
print("3. READING RESOURCES")
print("="*60)
# Read settings resource
print("\nFetching Calculator Settings...")
settings_resource = await client.read_resource("config://calculator/settings")
print(f" Version: {settings_resource[0].text}")
# Read guide resource
print("\nFetching Calculator Guide...")
guide_resource = await client.read_resource("docs://calculator/guide")
# Print first 200 characters of guide
guide_text = guide_resource[0].text[:200] + "..."
print(f" {guide_text}")
# CHAINING OPERATIONS
print("\n" + "="*60)
print("4. CHAINING MULTIPLE OPERATIONS")
print("="*60)
# Calculate: (10 + 5) * 3 - 7
print("\nCalculating: (10 + 5) * 3 - 7")
# Step 1: Add
print(" Step 1: Add 10 + 5")
add_result = await client.call_tool("add", {"a": 10, "b": 5})
step1 = extract_tool_result(add_result)
print(f" Result: {step1}")
# Step 2: Multiply
print(" Step 2: Multiply 15 * 3")
mult_result = await client.call_tool("multiply", {"a": step1, "b": 3})
step2 = extract_tool_result(mult_result)
print(f" Result: {step2}")
# Step 3: Subtract
print(" Step 3: Subtract 45 - 7")
final_result = await client.call_tool("subtract", {"a": step2, "b": 7})
final = extract_tool_result(final_result)
print(f" Final Result: {final}")
# GET PROMPT TEMPLATE
print("\n" + "="*60)
print("5. USING PROMPT TEMPLATES")
print("="*60)
expression = "25 * 4 + 10 / 2"
print(f"\nPrompt Template for: {expression}")
prompt_response = await client.get_prompt(
"calculate_expression",
{"expression": expression}
)
print(f" Template:\n{prompt_response.messages[0].content.text}")
logger.info("✓ Client operations completed successfully")
except Exception as e:
logger.error(f"Client error: {e}", exc_info=True)
sys.exit(1)
From the code above, the client uses async with Client(server) for safe connection management. This automatically handles connection setup and cleanup.
We also need a helper function to handle the results.
def extract_tool_result(response: Any) -> Any:
"""
Extract the actual result value from a tool response.
MCP wraps results in content objects, this helper unwraps them.
"""
try:
if hasattr(response, 'content') and response.content:
content = response.content[0]
# Prefer explicit text content when available (TextContent)
if hasattr(content, 'text') and content.text is not None:
# If the text is JSON, try to parse and extract a `result` field
import json as _json
text_val = content.text
try:
parsed_text = _json.loads(text_val)
# If JSON contains a result field, return it
if isinstance(parsed_text, dict) and 'result' in parsed_text:
return parsed_text.get('result')
return parsed_text
except _json.JSONDecodeError:
# Try to convert plain text to number
try:
if '.' in text_val:
return float(text_val)
return int(text_val)
except Exception:
return text_val
# Try to extract JSON result via model `.json()` or dict-like `.json`
if hasattr(content, 'json'):
try:
if callable(content.json):
json_str = content.json()
import json as _json
try:
parsed = _json.loads(json_str)
except _json.JSONDecodeError:
return json_str
else:
parsed = content.json
# If parsed is a dict, try common shapes
if isinstance(parsed, dict):
# If nested result exists
if 'result' in parsed:
res = parsed.get('result')
elif 'text' in parsed:
res = parsed.get('text')
else:
res = parsed
# If res is str that looks like a number, convert
if isinstance(res, str):
try:
if '.' in res:
return float(res)
return int(res)
except Exception:
return res
return res
return parsed
except Exception:
pass
return response
except Exception as e:
logger.warning(f"Could not extract result: {e}")
return response
if __name__ == "__main__":
logger.info("Calculator Client Starting...")
asyncio.run(main())
Looking at the above code, before using tools, the client lists available capabilities. The await client.list_tools() gets all tool metadata, including descriptions. The await client.list_resources() discovers available resources. Lastly, the await client.list_prompts() will find available prompt templates.
The await client.call_tool() method does the following:
- Takes the tool name and parameters as a dictionary
- Returns a wrapped response object containing the result
- Integrates with error handling for tool failures
On the result extraction, the extract_tool_result() helper function unwraps MCP’s response format to get the actual value, handling both JSON and text responses.
The chaining operations you see above demonstrate how to use output from one tool as input to another, enabling complex calculations across multiple tool calls.
Lastly, the error handling catches tool errors (like division by zero) and logs them gracefully without crashing.
// Step 4: Running the Server and Client
You will open two terminals. On terminal 1, you will start the server:
python calculator_server.py
You should see:
Image by Author
On terminal 2 run the client:
python calculator_client.py
Output will show:
Image by Author
# Advanced Patterns with FastMCP
While our calculator example uses basic logic, FastMCP is designed to handle complex, production-ready scenarios. As you scale your MCP servers, you can leverage:
- Asynchronous Operations: Use
async deffor tools that perform I/O-bound tasks like database queries or API calls - Dynamic Resources: Resources can accept arguments (e.g.,
resource://users/{user_id}) to fetch specific data points on the fly - Complex Type Validation: Use Pydantic models or complex Python type hints to ensure the LLM sends data in the exact format your backend requires
- Custom Transports: While we used
stdio, FastMCP also supports SSE (Server-Sent Events) for web-based integrations and custom UI tools
# Conclusion
FastMCP bridges the gap between the complex Model Context Protocol and the clean, decorator-based developer experience Python programmers expect. By removing the boilerplate associated with JSON-RPC 2.0 and manual transport management, it allows you to focus on what matters: building the tools that make LLMs more capable.
In this tutorial, we covered:
- The core architecture of MCP (Servers vs. Clients)
- How to define Tools for actions, Resources for data, and Prompts for instructions
- How to build a functional client to test and chain your server logic
Whether you are building a simple utility or a complex data orchestration layer, FastMCP provides the most “Pythonic” path to a production-ready agentic ecosystem.
What will you build next? Check out the FastMCP documentation to explore more advanced deployment strategies and UI integrations.
Shittu Olumide is a software engineer and technical writer passionate about leveraging cutting-edge technologies to craft compelling narratives, with a keen eye for detail and a knack for simplifying complex concepts. You can also find Shittu on Twitter.
