MCP Tool Server (mcp_server.py)

The mcp_server.py script is an asynchronous server designed to expose Python functions as "tools" over various communication transports. It utilizes the FastMCP framework to manage tools and handle client interactions. The server comes with built-in tools and supports dynamic loading of custom tools from Python files or directories.

Features

• Multiple Transports: Supports communication via:
  • stdio: Standard input/output, suitable for direct command-line interaction or integration with other processes.
  • sse: Server-Sent Events, allowing web clients to subscribe to tool outputs (listens on http://<host>:<port>/sse).
• Built-in Tools:
  • echo: Returns the parameters it was called with.
  • health: Reports the server's operational status (returns "ready").
• Custom Tool Loading:
  • Dynamically loads Python functions as tools from specified .py files or directories containing .py files.
  • Each tool file must define exactly one function to be exposed as a tool.
  • Imports within tool files are supported.
• Configurable Logging:
  • Outputs INFO level messages to the console by default.
  • Supports detailed DEBUG level logging to a specified file for both the server application and the underlying MCP SDK.
• Asynchronous: Built with asyncio for efficient handling of concurrent operations.

Prerequisites

• Python 3.x
• The mcp Python package (specifically mcp.server.fastmcp is used). Ensure this package is installed in your Python environment.

Command-Line Usage

Synopsis

python mcp_server.py [tools_paths...] [--log FILENAME] [--transport {stdio,sse}] [--port PORT]

Options

• tools_paths (positional, zero or more)

  • Paths to Python tool files or directories containing Python tool files.
  • If a directory is provided, the server will search for *.py files (excluding those starting with _).
  • Example: ./my_tool.py ./utils_directory
  • Default: None (only built-in tools will be available if no paths are provided).

• --log FILENAME

  • Enable detailed logging (DEBUG level for the application and MCP SDK) to the specified file.
  • Includes full details of tool calls, responses, and internal operations.
  • Console will still show INFO level messages.

• --transport {stdio|sse}

  • Communication transport to use.
  • stdio: Use standard input/output.
  • sse: Use Server-Sent Events. The server will listen on 0.0.0.0 by default.
  • Default: stdio

• --port PORT

  • Port number to use for the SSE transport.
  • Default: 8080

Tool Loading

Custom tools are Python functions that the server can expose.

• Definition: Each Python file (.py) intended to be a tool module should define exactly one function. This function will be registered as a tool. The name of the function will be its tool name.
• Discovery:
  • If a path to a .py file is provided, the server attempts to load the tool from that file.
  • If a path to a directory is provided, the server scans the directory for .py files (ignoring files like __init__.py or any file starting with an underscore) and attempts to load a tool from each valid file.
• Imports: Tool files can import other modules as needed. The server temporarily adds the tool's directory to sys.path during loading to facilitate local imports.

Examples

1. Creating a Custom Tool

Create a Python file, for example, calculator.py:

# calculator.py
def add(a: int, b: int) -> int:
    """
    Adds two integers and returns the result.
    """
    print(f"Tool 'add' called with a={a}, b={b}") # Example of internal tool logging
    return a + b

2. Running the Server

a) With built-in tools only (stdio transport):

python mcp_server.py

The server will start, and you can interact with echo and health tools via its standard input/output (details depend on the FastMCP client interaction protocol).

b) With a custom tool (stdio transport):

python mcp_server.py calculator.py

Now, the add tool from calculator.py will also be available.

c) With a directory of tools and detailed logging (stdio transport):

Assuming you have a directory named my_tools containing calculator.py and other tool files:

mkdir my_tools
mv calculator.py my_tools/
# You can add more .py tool files to my_tools/

python mcp_server.py my_tools --log server.log

The server will load all valid tools from the my_tools directory. Detailed logs will be written to server.log, and INFO messages to the console.

d) Using SSE transport:

python mcp_server.py calculator.py --transport sse --port 8888

The server will start and listen for SSE connections on http://0.0.0.0:8888/sse. Clients can connect to this endpoint to interact with the echo, health, and add tools.

Interacting with the Server (Conceptual)

How you interact with the tools depends on the chosen transport and the FastMCP client-side implementation.

• For stdio: You would typically send JSON-RPC like messages to the server's standard input and receive responses on its standard output.

  // Example conceptual input for stdio (actual format depends on FastMCP)
  // {"jsonrpc": "2.0", "method": "add", "params": {"a": 5, "b": 3}, "id": 1}
  

• For sse: Clients would connect to the /sse endpoint. The server would then push messages for tool invocations and responses.

Logging

• Console: By default, INFO level messages are printed to stdout. This includes server startup messages, registered tools, and basic activity.
• File Logging (--log option):
  • Enables DEBUG level logging for both mcp_server.py and the mcp SDK.
  • This is highly verbose and includes:
    • Detailed tool invocation parameters and results.
    • Internal server operations.
    • Debug messages from the MCP SDK.
  • The log file is written in a format like: %(asctime)s - %(name)s - %(levelname)s - %(message)s.

Shutdown

Press Ctrl+C to shut down the server gracefully.