Building Production MCP (Model Context Protocol) Servers in Python for Claude Code & Cursor

Anthropic's Model Context Protocol (MCP) has established itself as the open universal standard connecting AI coding assistants (Claude Code, Cursor, Windsurf) with real-world infrastructure, local databases, and execution runtimes. Here is how to engineer production-ready FastMCP servers in Python without crashing your JSON-RPC connection.

1. The Architecture of FastMCP

FastMCP simplifies server creation by eliminating boilerplate JSON-RPC protocol handlers. Using straightforward Python decorators, you expose native functions directly as LLM tools with strict Pydantic type safety.

from mcp.server.fastmcp import FastMCP
import psycopg2, os

mcp = FastMCP("postgres-inspector")

@mcp.tool()
def inspect_schema(table_name: str) -> str:
    """Introspect PostgreSQL table schema and return column definitions safely."""
    # Read-only query execution with bounded output
    return query_db_schema_safely(table_name)

if __name__ == "__main__":
    mcp.run()

2. Solving Stdio Pollution & 'Connection Closed' Errors

The #1 error developers encounter when running local MCP servers is the fatal 'Connection closed' crash. This is caused by stray print() statements or third-party libraries writing debug output to standard output (stdout), which corrupts the JSON-RPC message frame.

• Stderr Isolation: Redirect all application logs exclusively to sys.stderr or dedicated log files on disk.
• Virtualenv PATH Resolution: Explicitly specify the absolute path to your Python interpreter inside claude_desktop_config.json.
• Graceful Shutdown: Implement SIGINT and SIGTERM traps to release database connection pools cleanly.