Context

Model Context Protocols (MCPs) are coming online because of AI

What is it?

It’s a standard dictates the format for defining:

Resources

Passive data sources defined like such:

{
  "uri": "nginx://logs/error",
  "name": "Nginx Error Logs",
  "mimeType": "text/plain",
  "description": "Live stream of the latest Nginx error logs"
}

Tools

Executable Functions. You are providing actions that the agent can take.

{
  "name": "update_nginx_map",
  "description": "Updates the IP allowlist in the Nginx config",
  "inputSchema": {
    "type": "object",
    "properties": {
      "ip_address": { "type": "string" },
      "action": { "type": "integer", "enum": [0, 1] }
    },
    "required": ["ip_address", "action"]
  }
}

Prompts

Workflows, templates, Instructions. You are providing task lists the agent can take.

{
  "name": "debug-gateway",
  "description": "Start a step-by-step diagnostic of the Nginx gateway",
  "arguments": [
    {
      "name": "service_name",
      "description": "The Go service to check",
      "required": true
    }
  ]
}

File/Folder Structure

The files are organized as such:

/nginx-mcp-server
├── server.py          <-- The "Brain" (Where decorators live)
├── pyproject.toml     <-- Dependencies (mcp[cli], httpx, etc.)
└── .env               <-- Secret keys (if your API needs them)

FastMCP

It providers you a way to write MCP classes using Python instead of writing json. It includes logging to client, progress reporting, hot reloading.

Handshake

The LLM client (VSCode or Claude Desktop) reads the local config file. The mcp.json file.

{
  "mcpServers": {
    "nginx-manager": {
      "command": "python",
      "args": ["/path/to/server.py"]
    }
  }
}

A list of mcpServers

The client then performs a handshake. The JSON-RPC Discovery phase. The Python code is converted into the JSON-RPC Schema at this time.
The client receives the schema and processes it for use.