mcp-template

MCP template.

Development

TODOs

Find # TODO comments and implement them.

Logging

Do not use print statements for logging. Use the logging module instead. Writing to stdout will corrupt the JSON-RPC messages and break your server.

Docstrings / Tool decorator parameters

MCP.tools decorator parameters are especially important as this is the human readable text that the LLM has context of. This will be treated as part of the prompt when fed to the LLM and this will decide when to use each tool.

Installation

Option 1: Development setup with uv

Get repo:

git clone https://github.com/xcollantes/mcp-template.git  # TODO: Change to your own repository.
cd mcp-template

Install UV: https://docs.astral.sh/uv/getting-started/installation/

Option 2: Install globally with pipx

# Install pipx if you haven't already
brew install pipx
pipx ensurepath

# Clone and install the MCP server
git clone https://github.com/xcollantes/mcp-template.git  # TODO: Change to your own repository.
cd mcp-template
pipx install -e .

Now you can run the MCP server directly:

mcp-template --help
mcp-template --debug  # Run with debug logging

Option 3: Manual setup

Add MCP server to your choice of LLM client:

NOTE: You will need to look up for your specific client on how to add MCPs.

Usually the JSON file for the LLM client will look like this:

{
  "mcpServers": {
    "weather": {
      "command": "uv",
      "args": ["--directory", "/ABSOLUTE/PATH/TO/REPO", "run", "python", "-m", "src.main"]
    }
  }
}

This will tell your LLM client application that there's a tool that can be called by calling uv --directory /ABSOLUTE/PATH/TO/REPO run python -m src.main.

How it works

1. You enter some questions or prompt to a LLM Client such as the Claude Desktop, Cursor, Windsurf, or ChatGPT.
2. The client sends your question to the LLM model (Sonnet, Grok, ChatGPT)
3. LLM analyzes the available tools and decides which one(s) to use
   • The LLM you're using will have a context of the tools and what each tool is meant for in human language.
   • Alternatively without MCPs, you could include in the prompt the endpoints and a description on each endpoint for the LLM to "call on". Then you could copy and paste the text commands into the terminal on your machine.
   • MCPs provide a more deterministic and standardized method on LLM-to-server interactions.
4. The client executes the chosen tool(s) through the MCP server.
   • The MCP server is either running local on your machine or an endpoint hosting the MCP server remotely.
5. The results are sent back to LLM.
6. LLM formulates a natural language response and one or both of the following happen:
   • The response is displayed to you with data from the MCP server
   • Some action is performed using the MCP server

Architecture

MCP follows a client-server architecture where an MCP host (an AI application like Cursor or ChatGPT desktop) establishes connections to one or more MCP servers. The MCP host accomplishes this by creating one MCP client for each MCP server. Each MCP client maintains a dedicated connection with its corresponding MCP server.

https://modelcontextprotocol.io/docs/learn/architecture