Vidhan Solanki | Portfolio

How actually Claude or Chatgpt are using the MCP tool I created this weekend?

LLMs usually follow a very specific "handshake" process

Here is exactly what is happening

1. The Discovery Phase

When you first add the URL to ChatGPT or start a session, ChatGPT sends a POST request to your /mcp endpoint.

ChatGPT asks: "What can you do?"
Your Server responds: "I am a server called weathervyrs. I have two tools: get_alerts (which needs a state code) and get_forecast (which needs latitude and longitude)."
Crucial part: Your server doesn't just send the names; it sends the Zod schema you defined. This tells the LLM exactly what types of data (strings, numbers, ranges) it is allowed to send.

2. The Decision Phase

When you type "What are the active alerts for NY", the LLM doesn't just search the web. It looks at its available "tools"

It sees that your get_alerts tool takes a state string.
It knows that "New York" matches the state NY.
The LLM decides to call your tool.

The Execution Phase (The JSON-RPC Request)

The LLM sends another POST request to your /mcp endpoint. This request is formatted in JSON-RPC. It looks something like this:

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "get_alerts",
    "arguments": { "state": "NY" }
  },
  "id": "123"
}

Your server receives this, your TypeScript code runs the makeNWSRequest function, fetches the real data from the government weather API, and sends it back as text

3. The Synthesis Phase

The LLM receives the raw text (like "Active alerts for NY: Winter Storm Warning..."). It reads that data, summarizes it, and types out a friendly response to you in the chat.

Hurdles I got

hard to connect via claude.ai (the tutorial used claude desktop which was not available on linux rn and I also didn't bother to check any alternative) so I decide to use ngrok in order to get a working mcp out for testing on the claude website directly Remembering points