Code beginner · 3 min read

How to build an MCP server in Python

Direct answer

Use the official mcp Python SDK to build an MCP server with stdio_server from mcp.server.stdio, which handles communication over standard input/output streams.

Setup

Install

bash

pip install mcp

Imports

python

from mcp.server import Server
from mcp.server.stdio import stdio_server

Examples

inStart the MCP server and wait for agent requests.

outMCP server started, listening on stdio for requests...

inReceive a JSON request from an AI agent to call a tool.

outReceived request: {'tool': 'calculator', 'input': '2+2'} Responded with: {'result': 4}

inHandle a malformed request gracefully.

outError: Invalid request format, sending error response.

Integration steps

Install the official MCP Python SDK with pip.
Import Server and stdio_server from mcp.server and mcp.server.stdio respectively.
Define your tool handlers or agent logic as functions or classes.
Create an MCP Server instance and register your tools.
Run stdio_server(server) to start the server listening on standard input/output.
Process incoming requests and send responses via the MCP protocol automatically.

Full code

python

from mcp.server import Server
from mcp.server.stdio import stdio_server

# Define a simple tool handler
class CalculatorTool:
    def call(self, input):
        try:
            # Evaluate simple math expressions safely
            result = eval(input, {"__builtins__": None}, {})
            return {"result": result}
        except Exception as e:
            return {"error": str(e)}

# Create MCP server instance
server = Server()

# Register the tool under a name
server.register_tool("calculator", CalculatorTool())

# Start the MCP server using stdio transport
print("Starting MCP server on stdio...")
stdio_server(server)

output

Starting MCP server on stdio...
# (Server now waits for JSON requests on stdin and responds on stdout)

API trace

Request

json

{"tool": "calculator", "input": "2+2"}

Response

json

{"result": 4}

Extractresponse['result'] or response.get('error')

Variants

Async MCP Server with stdio ›

Use when your tool handlers require asynchronous operations or I/O.

python

import asyncio
from mcp.server import Server
from mcp.server.stdio import stdio_server

class AsyncCalculatorTool:
    async def call(self, input):
        # Async evaluation or external API call
        return {"result": eval(input, {"__builtins__": None}, {})}

server = Server()
server.register_tool("calculator", AsyncCalculatorTool())

async def main():
    print("Starting async MCP server on stdio...")
    await stdio_server(server)

asyncio.run(main())

MCP Server with SSE Transport ›

Use SSE transport for web-based or networked MCP server deployments instead of stdio.

python

# SSE transport example is conceptual; requires custom SSE server setup
from mcp.server import Server
from mcp.server.sse import sse_server

server = Server()
# Register tools as usual

print("Starting MCP server on SSE transport...")
sse_server(server, host='0.0.0.0', port=8080)

Performance

Latency~10-50ms per request depending on tool complexity

CostFree, runs locally with no API cost

Rate limitsNone, limited only by local system resources

Keep tool input concise to reduce parsing overhead.
Avoid sending large payloads over stdio to maintain responsiveness.

Approach	Latency	Cost/call	Best for
Synchronous stdio MCP server	~10-50ms	Free	Simple local tool integration
Asynchronous stdio MCP server	~10-100ms	Free	Tools requiring async I/O
SSE transport MCP server	~20-100ms	Free	Networked or web-based MCP servers

✓

Quick tip

Always register your tools with clear, unique names on the MCP Server instance before starting the server.

⚠

Common mistake

Beginners often forget to run the server with stdio_server(server), causing no communication over stdio.

Verified 2026-04

Verify ↗