Python implementation of the Model Context Protocol (MCP)
- MCP Python SDK
The Model Context Protocol allows applications to provide context for LLMs in a standardized way, separating the concerns of providing context from the actual LLM interaction. This Python SDK implements the full MCP specification, making it easy to:
- Build MCP clients that can connect to any MCP server
- Create MCP servers that expose resources, prompts and tools
- Use standard transports like stdio, SSE, and Streamable HTTP
- Handle all MCP protocol messages and lifecycle events
We recommend using uv to manage your Python projects.
If you haven't created a uv-managed project yet, create one:
uv init mcp-server-demo
cd mcp-server-demoThen add MCP to your project dependencies:
uv add "mcp[cli]"Alternatively, for projects using pip for dependencies:
pip install "mcp[cli]"To run the mcp command with uv:
uv run mcpLet's create a simple MCP server that exposes a calculator tool and some data:
# server.py
from mcp.server.fastmcp import FastMCP
# Create an MCP server
mcp = FastMCP("Demo")
# Add an addition tool
@mcp.tool()
def add(a: int, b: int) -> int:
"""Add two numbers"""
return a + b
# Add a dynamic greeting resource
@mcp.resource("greeting://{name}")
def get_greeting(name: str) -> str:
"""Get a personalized greeting"""
return f"Hello, {name}!"You can install this server in Claude Desktop and interact with it right away by running:
mcp install server.pyAlternatively, you can test it with the MCP Inspector:
mcp dev server.pyThe Model Context Protocol (MCP) lets you build servers that expose data and functionality to LLM applications in a secure, standardized way. Think of it like a web API, but specifically designed for LLM interactions. MCP servers can:
- Expose data through Resources (think of these sort of like GET endpoints; they are used to load information into the LLM's context)
- Provide functionality through Tools (sort of like POST endpoints; they are used to execute code or otherwise produce a side effect)
- Define interaction patterns through Prompts (reusable templates for LLM interactions)
- And more!
The FastMCP server is your core interface to the MCP protocol. It handles connection management, protocol compliance, and message routing:
# Add lifespan support for startup/shutdown with strong typing
from contextlib import asynccontextmanager
from collections.abc import AsyncIterator
from dataclasses import dataclass
from fake_database import Database # Replace with your actual DB type
from mcp.server.fastmcp import FastMCP
# Create a named server
mcp = FastMCP("My App")
# Specify dependencies for deployment and development
mcp = FastMCP("My App", dependencies=["pandas", "numpy"])
@dataclass
class AppContext:
db: Database
@asynccontextmanager
async def app_lifespan(server: FastMCP) -> AsyncIterator[AppContext]:
"""Manage application lifecycle with type-safe context"""
# Initialize on startup
db = await Database.connect()
try:
yield AppContext(db=db)
finally:
# Cleanup on shutdown
await db.disconnect()
# Pass lifespan to server
mcp = FastMCP("My App", lifespan=app_lifespan)
# Access type-safe lifespan context in tools
@mcp.tool()
def query_db() -> str:
"""Tool that uses initialized resources"""
ctx = mcp.get_context()
db = ctx.request_context.lifespan_context["db"]
return db.query()Resources are how you expose data to LLMs. They're similar to GET endpoints in a REST API - they provide data but shouldn't perform significant computation or have side effects:
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("My App")
@mcp.resource("config://app", title="Application Configuration")
def get_config() -> str:
"""Static configuration data"""
return "App configuration here"
@mcp.resource("users://{user_id}/profile", title="User Profile")
def get_user_profile(user_id: str) -> str:
"""Dynamic user data"""
return f"Profile data for user {user_id}"Tools let LLMs take actions through your server. Unlike resources, tools are expected to perform computation and have side effects:
import httpx
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("My App")
@mcp.tool(title="BMI Calculator")
def calculate_bmi(weight_kg: float, height_m: float) -> float:
"""Calculate BMI given weight in kg and height in meters"""
return weight_kg / (height_m**2)
@mcp.tool(title="Weather Fetcher")
async def fetch_weather(city: str) -> str:
"""Fetch current weather for a city"""
async with httpx.AsyncClient() as client:
response = await client.get(f"https://api.weather.com/{city}")
return response.textPrompts are reusable templates that help LLMs interact with your server effectively:
from mcp.server.fastmcp import FastMCP
from mcp.server.fastmcp.prompts import base
mcp = FastMCP("My App")
@mcp.prompt(title="Code Review")
def review_code(code: str) -> str:
return f"Please review this code:\n\n{code}"
@mcp.prompt(title="Debug Assistant")
def debug_error(error: str) -> list[base.Message]:
return [
base.UserMessage("I'm seeing this error:"),
base.UserMessage(error),
base.AssistantMessage("I'll help debug that. What have you tried so far?"),
]FastMCP provides an Image class that automatically handles image data:
from mcp.server.fastmcp import FastMCP, Image
from PIL import Image as PILImage
mcp = FastMCP("My App")
@mcp.tool()
def create_thumbnail(image_path: str) -> Image:
"""Create a thumbnail from an image"""
img = PILImage.open(image_path)
img.thumbnail((100, 100))
return Image(data=img.tobytes(), format="png")The Context object gives your tools and resources access to MCP capabilities:
from mcp.server.fastmcp import FastMCP, Context
mcp = FastMCP("My App")
@mcp.tool()
async def long_task(files: list[str], ctx: Context) -> str:
"""Process multiple files with progress tracking"""
for i, file in enumerate(files):
ctx.info(f"Processing {file}")
await ctx.report_progress(i, len(files))
data, mime_type = await ctx.read_resource(f"file://{file}")
return "Processing complete"MCP supports providing completion suggestions for prompt arguments and resource template parameters. With the context parameter, servers can provide completions based on previously resolved values:
Client usage:
from mcp.client.session import ClientSession
from mcp.types import ResourceTemplateReference
async def use_completion(session: ClientSession):
# Complete without context
result = await session.complete(
ref=ResourceTemplateReference(
type="ref/resource", uri="github://repos/{owner}/{repo}"
),
argument={"name": "owner", "value": "model"},
)
# Complete with context - repo suggestions based on owner
result = await session.complete(
ref=ResourceTemplateReference(
type="ref/resource", uri="github://repos/{owner}/{repo}"
),
argument={"name": "repo", "value": "test"},
context_arguments={"owner": "modelcontextprotocol"},
)Server implementation:
from mcp.server import Server
from mcp.types import (
Completion,
CompletionArgument,
CompletionContext,
PromptReference,
ResourceTemplateReference,
)
server = Server("example-server")
@server.completion()
async def handle_completion(
ref: PromptReference | ResourceTemplateReference,
argument: CompletionArgument,
context: CompletionContext | None,
) -> Completion | None:
if isinstance(ref, ResourceTemplateReference):
if ref.uri == "github://repos/{owner}/{repo}" and argument.name == "repo":
# Use context to provide owner-specific repos
if context and context.arguments:
owner = context.arguments.get("owner")
if owner == "modelcontextprotocol":
repos = ["python-sdk", "typescript-sdk", "specification"]
# Filter based on partial input
filtered = [r for r in repos if r.startswith(argument.value)]
return Completion(values=filtered)
return NoneRequest additional information from users during tool execution:
from mcp.server.fastmcp import FastMCP, Context
from mcp.server.elicitation import (
AcceptedElicitation,
DeclinedElicitation,
CancelledElicitation,
)
from pydantic import BaseModel, Field
mcp = FastMCP("Booking System")
@mcp.tool()
async def book_table(date: str, party_size: int, ctx: Context) -> str:
"""Book a table with confirmation"""
# Schema must only contain primitive types (str, int, float, bool)
class ConfirmBooking(BaseModel):
confirm: bool = Field(description="Confirm booking?")
notes: str = Field(default="", description="Special requests")
result = await ctx.elicit(
message=f"Confirm booking for {party_size} on {date}?", schema=ConfirmBooking
)
match result:
case AcceptedElicitation(data=data):
if data.confirm:
return f"Booked! Notes: {data.notes or 'None'}"
return "Booking cancelled"
case DeclinedElicitation():
return "Booking declined"
case CancelledElicitation():
return "Booking cancelled"The elicit() method returns an ElicitationResult with:
action: "accept", "decline", or "cancel"data: The validated response (only when accepted)validation_error: Any validation error message
Authentication can be used by servers that want to expose tools accessing protected resources.
mcp.server.auth implements an OAuth 2.0 server interface, which servers can use by
providing an implementation of the OAuthAuthorizationServerProvider protocol.
from mcp import FastMCP
from mcp.server.auth.provider import OAuthAuthorizationServerProvider
from mcp.server.auth.settings import (
AuthSettings,
ClientRegistrationOptions,
RevocationOptions,
)
class MyOAuthServerProvider(OAuthAuthorizationServerProvider):
# See an example on how to implement at `examples/servers/simple-auth`
...
mcp = FastMCP(
"My App",
auth_server_provider=MyOAuthServerProvider(),
auth=AuthSettings(
issuer_url="https://myapp.com",
revocation_options=RevocationOptions(
enabled=True,
),
client_registration_options=ClientRegistrationOptions(
enabled=True,
valid_scopes=["myscope", "myotherscope"],
default_scopes=["myscope"],
),
required_scopes=["myscope"],
),
)See OAuthAuthorizationServerProvider for more details.
The fastest way to test and debug your server is with the MCP Inspector:
mcp dev server.py
# Add dependencies
mcp dev server.py --with pandas --with numpy
# Mount local code
mcp dev server.py --with-editable .Once your server is ready, install it in Claude Desktop:
mcp install server.py
# Custom name
mcp install server.py --name "My Analytics Server"
# Environment variables
mcp install server.py -v API_KEY=abc123 -v DB_URL=postgres://...
mcp install server.py -f .envFor advanced scenarios like custom deployments:
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("My App")
if __name__ == "__main__":
mcp.run()Run it with:
python server.py
# or
mcp run server.pyNote that mcp run or mcp dev only supports server using FastMCP and not the low-level server variant.
Note: Streamable HTTP transport is superseding SSE transport for production deployments.
from mcp.server.fastmcp import FastMCP
# Stateful server (maintains session state)
mcp = FastMCP("StatefulServer")
# Stateless server (no session persistence)
mcp = FastMCP("StatelessServer", stateless_http=True)
# Stateless server (no session persistence, no sse stream with supported client)
mcp = FastMCP("StatelessServer", stateless_http=True, json_response=True)
# Run server with streamable_http transport
mcp.run(transport="streamable-http")You can mount multiple FastMCP servers in a FastAPI application:
# echo.py
from mcp.server.fastmcp import FastMCP
mcp = FastMCP(name="EchoServer", stateless_http=True)
@mcp.tool(description="A simple echo tool")
def echo(message: str) -> str:
return f"Echo: {message}"# math.py
from mcp.server.fastmcp import FastMCP
mcp = FastMCP(name="MathServer", stateless_http=True)
@mcp.tool(description="A simple add tool")
def add_two(n: int) -> int:
return n + 2# main.py
import contextlib
from fastapi import FastAPI
from mcp.echo import echo
from mcp.math import math
# Create a combined lifespan to manage both session managers
@contextlib.asynccontextmanager
async def lifespan(app: FastAPI):
async with contextlib.AsyncExitStack() as stack:
await stack.enter_async_context(echo.mcp.session_manager.run())
await stack.enter_async_context(math.mcp.session_manager.run())
yield
app = FastAPI(lifespan=lifespan)
app.mount("/echo", echo.mcp.streamable_http_app())
app.mount("/math", math.mcp.streamable_http_app())For low level server with Streamable HTTP implementations, see:
- Stateful server:
examples/servers/simple-streamablehttp/ - Stateless server:
examples/servers/simple-streamablehttp-stateless/
The streamable HTTP transport supports:
- Stateful and stateless operation modes
- Resumability with event stores
- JSON or SSE response formats
- Better scalability for multi-node deployments
Note: SSE transport is being superseded by Streamable HTTP transport.
By default, SSE servers are mounted at /sse and Streamable HTTP servers are mounted at /mcp. You can customize these paths using the methods described below.
You can mount the SSE server to an existing ASGI server using the sse_app method. This allows you to integrate the SSE server with other ASGI applications.
from starlette.applications import Starlette
from starlette.routing import Mount, Host
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("My App")
# Mount the SSE server to the existing ASGI server
app = Starlette(
routes=[
Mount('/', app=mcp.sse_app()),
]
)
# or dynamically mount as host
app.router.routes.append(Host('mcp.acme.corp', app=mcp.sse_app()))When mounting multiple MCP servers under different paths, you can configure the mount path in several ways:
from starlette.applications import Starlette
from starlette.routing import Mount
from mcp.server.fastmcp import FastMCP
# Create multiple MCP servers
github_mcp = FastMCP("GitHub API")
browser_mcp = FastMCP("Browser")
curl_mcp = FastMCP("Curl")
search_mcp = FastMCP("Search")
# Method 1: Configure mount paths via settings (recommended for persistent configuration)
github_mcp.settings.mount_path = "/github"
browser_mcp.settings.mount_path = "/browser"
# Method 2: Pass mount path directly to sse_app (preferred for ad-hoc mounting)
# This approach doesn't modify the server's settings permanently
# Create Starlette app with multiple mounted servers
app = Starlette(
routes=[
# Using settings-based configuration
Mount("/github", app=github_mcp.sse_app()),
Mount("/browser", app=browser_mcp.sse_app()),
# Using direct mount path parameter
Mount("/curl", app=curl_mcp.sse_app("/curl")),
Mount("/search", app=search_mcp.sse_app("/search")),
]
)
# Method 3: For direct execution, you can also pass the mount path to run()
if __name__ == "__main__":
search_mcp.run(transport="sse", mount_path="/search")For more information on mounting applications in Starlette, see the Starlette documentation.
A simple server demonstrating resources, tools, and prompts:
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("Echo")
@mcp.resource("echo://{message}")
def echo_resource(message: str) -> str:
"""Echo a message as a resource"""
return f"Resource echo: {message}"
@mcp.tool()
def echo_tool(message: str) -> str:
"""Echo a message as a tool"""
return f"Tool echo: {message}"
@mcp.prompt()
def echo_prompt(message: str) -> str:
"""Create an echo prompt"""
return f"Please process this message: {message}"A more complex example showing database integration:
import sqlite3
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("SQLite Explorer")
@mcp.resource("schema://main")
def get_schema() -> str:
"""Provide the database schema as a resource"""
conn = sqlite3.connect("database.db")
schema = conn.execute("SELECT sql FROM sqlite_master WHERE type='table'").fetchall()
return "\n".join(sql[0] for sql in schema if sql[0])
@mcp.tool()
def query_data(sql: str) -> str:
"""Execute SQL queries safely"""
conn = sqlite3.connect("database.db")
try:
result = conn.execute(sql).fetchall()
return "\n".join(str(row) for row in result)
except Exception as e:
return f"Error: {str(e)}"For more control, you can use the low-level server implementation directly. This gives you full access to the protocol and allows you to customize every aspect of your server, including lifecycle management through the lifespan API:
from contextlib import asynccontextmanager
from collections.abc import AsyncIterator
from fake_database import Database # Replace with your actual DB type
from mcp.server import Server
@asynccontextmanager
async def server_lifespan(server: Server) -> AsyncIterator[dict]:
"""Manage server startup and shutdown lifecycle."""
# Initialize resources on startup
db = await Database.connect()
try:
yield {"db": db}
finally:
# Clean up on shutdown
await db.disconnect()
# Pass lifespan to server
server = Server("example-server", lifespan=server_lifespan)
# Access lifespan context in handlers
@server.call_tool()
async def query_db(name: str, arguments: dict) -> list:
ctx = server.request_context
db = ctx.lifespan_context["db"]
return await db.query(arguments["query"])The lifespan API provides:
- A way to initialize resources when the server starts and clean them up when it stops
- Access to initialized resources through the request context in handlers
- Type-safe context passing between lifespan and request handlers
import mcp.server.stdio
import mcp.types as types
from mcp.server.lowlevel import NotificationOptions, Server
from mcp.server.models import InitializationOptions
# Create a server instance
server = Server("example-server")
@server.list_prompts()
async def handle_list_prompts() -> list[types.Prompt]:
return [
types.Prompt(
name="example-prompt",
description="An example prompt template",
arguments=[
types.PromptArgument(
name="arg1", description="Example argument", required=True
)
],
)
]
@server.get_prompt()
async def handle_get_prompt(
name: str, arguments: dict[str, str] | None
) -> types.GetPromptResult:
if name != "example-prompt":
raise ValueError(f"Unknown prompt: {name}")
return types.GetPromptResult(
description="Example prompt",
messages=[
types.PromptMessage(
role="user",
content=types.TextContent(type="text", text="Example prompt text"),
)
],
)
async def run():
async with mcp.server.stdio.stdio_server() as (read_stream, write_stream):
await server.run(
read_stream,
write_stream,
InitializationOptions(
server_name="example",
server_version="0.1.0",
capabilities=server.get_capabilities(
notification_options=NotificationOptions(),
experimental_capabilities={},
),
),
)
if __name__ == "__main__":
import asyncio
asyncio.run(run())Caution: The mcp run and mcp dev tool doesn't support low-level server.
The SDK provides a high-level client interface for connecting to MCP servers using various transports:
from mcp import ClientSession, StdioServerParameters, types
from mcp.client.stdio import stdio_client
# Create server parameters for stdio connection
server_params = StdioServerParameters(
command="python", # Executable
args=["example_server.py"], # Optional command line arguments
env=None, # Optional environment variables
)
# Optional: create a sampling callback
async def handle_sampling_message(
message: types.CreateMessageRequestParams,
) -> types.CreateMessageResult:
return types.CreateMessageResult(
role="assistant",
content=types.TextContent(
type="text",
text="Hello, world! from model",
),
model="gpt-3.5-turbo",
stopReason="endTurn",
)
async def run():
async with stdio_client(server_params) as (read, write):
async with ClientSession(
read, write, sampling_callback=handle_sampling_message
) as session:
# Initialize the connection
await session.initialize()
# List available prompts
prompts = await session.list_prompts()
# Get a prompt
prompt = await session.get_prompt(
"example-prompt", arguments={"arg1": "value"}
)
# List available resources
resources = await session.list_resources()
# List available tools
tools = await session.list_tools()
# Read a resource
content, mime_type = await session.read_resource("file://some/path")
# Call a tool
result = await session.call_tool("tool-name", arguments={"arg1": "value"})
if __name__ == "__main__":
import asyncio
asyncio.run(run())Clients can also connect using Streamable HTTP transport:
from mcp.client.streamable_http import streamablehttp_client
from mcp import ClientSession
async def main():
# Connect to a streamable HTTP server
async with streamablehttp_client("example/mcp") as (
read_stream,
write_stream,
_,
):
# Create a session using the client streams
async with ClientSession(read_stream, write_stream) as session:
# Initialize the connection
await session.initialize()
# Call a tool
tool_result = await session.call_tool("echo", {"message": "hello"})When building MCP clients, the SDK provides utilities to help display human-readable names for tools, resources, and prompts:
from mcp.shared.metadata_utils import get_display_name
from mcp.client.session import ClientSession
async def display_tools(session: ClientSession):
"""Display available tools with human-readable names"""
tools_response = await session.list_tools()
for tool in tools_response.tools:
# get_display_name() returns the title if available, otherwise the name
display_name = get_display_name(tool)
print(f"Tool: {display_name}")
if tool.description:
print(f" {tool.description}")
async def display_resources(session: ClientSession):
"""Display available resources with human-readable names"""
resources_response = await session.list_resources()
for resource in resources_response.resources:
display_name = get_display_name(resource)
print(f"Resource: {display_name} ({resource.uri})")The get_display_name() function implements the proper precedence rules for displaying names:
- For tools:
title>annotations.title>name - For other objects:
title>name
This ensures your client UI shows the most user-friendly names that servers provide.
The SDK includes authorization support for connecting to protected MCP servers:
from mcp.client.auth import OAuthClientProvider, TokenStorage
from mcp.client.session import ClientSession
from mcp.client.streamable_http import streamablehttp_client
from mcp.shared.auth import OAuthClientInformationFull, OAuthClientMetadata, OAuthToken
class CustomTokenStorage(TokenStorage):
"""Simple in-memory token storage implementation."""
async def get_tokens(self) -> OAuthToken | None:
pass
async def set_tokens(self, tokens: OAuthToken) -> None:
pass
async def get_client_info(self) -> OAuthClientInformationFull | None:
pass
async def set_client_info(self, client_info: OAuthClientInformationFull) -> None:
pass
async def main():
# Set up OAuth authentication
oauth_auth = OAuthClientProvider(
server_url="https://api.example.com",
client_metadata=OAuthClientMetadata(
client_name="My Client",
redirect_uris=["http://localhost:3000/callback"],
grant_types=["authorization_code", "refresh_token"],
response_types=["code"],
),
storage=CustomTokenStorage(),
redirect_handler=lambda url: print(f"Visit: {url}"),
callback_handler=lambda: ("auth_code", None),
)
# Use with streamable HTTP client
async with streamablehttp_client(
"https://api.example.com/mcp", auth=oauth_auth
) as (read, write, _):
async with ClientSession(read, write) as session:
await session.initialize()
# Authenticated session readyFor a complete working example, see examples/clients/simple-auth-client/.
The MCP protocol defines three core primitives that servers can implement:
| Primitive | Control | Description | Example Use |
|---|---|---|---|
| Prompts | User-controlled | Interactive templates invoked by user choice | Slash commands, menu options |
| Resources | Application-controlled | Contextual data managed by the client application | File contents, API responses |
| Tools | Model-controlled | Functions exposed to the LLM to take actions | API calls, data updates |
MCP servers declare capabilities during initialization:
| Capability | Feature Flag | Description |
|---|---|---|
prompts |
listChanged |
Prompt template management |
resources |
subscribelistChanged |
Resource exposure and updates |
tools |
listChanged |
Tool discovery and execution |
logging |
- | Server logging configuration |
completion |
- | Argument completion suggestions |
- Model Context Protocol documentation
- Model Context Protocol specification
- Officially supported servers
We are passionate about supporting contributors of all levels of experience and would love to see you get involved in the project. See the contributing guide to get started.
This project is licensed under the MIT License - see the LICENSE file for details.