Request for clarification: Direct Execution (stdio transport) implementation

[grahama1970]

Overview

While exploring the MCP Python SDK, I noticed some differences between the documented direct execution approach and actual behavior. Looking for guidance or clarification.

Current Understanding

The documentation describes direct execution without npm:

from mcp.server.fastmcp import FastMCP
mcp = FastMCP("My App")
if __name__ == "__main__":
    mcp.run()

Testing Results

1. Web UI Implementation (Working):

mcp dev quickstart.py

This works perfectly - tools register and execute as expected.

2. Direct Execution Attempts:

# demo_server.py
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("Demo")

@mcp.tool()
def add(a: int, b: int) -> int:
    return a + b

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

Tried running with:

python demo_server.py
# and
(sleep 3; echo '{"jsonrpc": "2.0", "method": "tools/call", "params": {"name": "add", "args": {"a": 2, "b": 3}}, "id": 1}'; cat) | python demo_server.py

Questions

1. Is there additional setup needed for direct execution?
2. Are there working examples of stdio transport implementation?
3. Should the documentation be updated to clarify requirements?

Environment

• Python 3.11.6
• Latest MCP SDK
• macOS

Happy to provide more details or test cases if needed.