How MCP Tools Work • Jenny Yang‘s Personal Blog

MCP (Model Context Protocol) defines a standard way for AI agents to discover and invoke tools. Instead of hardcoding logic, the agent reads a tool definition at runtime and figures out how to call it based on the user’s intent.

The Flow#

The diagram below shows the full cycle for a simple weather query.

1. User sends a natural language query#

"What's the weather like in Shanghai today? Show me in Celsius."

plaintext

2. Agent decomposes intent and extracts parameters#

The agent matches the query against available tool descriptions:

Step	Result
intent match	`"weather"` → matches description of `get_weather`
city extract	`"Shanghai"` → `city = "Shanghai"`
unit extract	`"Celsius"` → `unit = "celsius"`

3. Agent reads the tool definition#

The tool definition tells the agent exactly what it needs to call the tool:

{
  "name": "get_weather",
  "description": "Get current weather and temperature for a given city",
  "parameters": {
    "city": "string - name of the city (required)",
    "unit": "string - \"celsius\" or \"fahrenheit\" (optional)"
  }
}

json

4. Agent generates and executes the tool call#

{
  "tool_name": "get_weather",
  "arguments": { "city": "Shanghai", "unit": "celsius" }
}

json

The tool runs, returns the result, and the agent formats it into a natural language response.

Why This Matters#

The agent never needs to know the implementation of get_weather. It only needs:

A description good enough to match intent
A parameter schema to know what to extract

This separation is what makes MCP tools composable — you can add new tools without changing the agent.