An MCP primer - AI communication standards

by Rap Payne

Model Context Protocol (MCP) has been described as the USB for AI. I like that. It’s an open standard that lets large language models (LLMs) talk to tools and to each other. Despite its newness, MCP is the defacto standard. You should definitely use it.

What kinds of tools can MCP connect to?

Remember, your LLM was trained on data as of a certain date. It doesn’t know anything that happened after that date. But with MCP, it can call out to tools that do know about the world. For example:

Why have it?

If we didn’t have MCP, then every LLM would have a different way of talking. That would be horrible for devs, who’d have to re-learn how to communicate with every new model. But with MCP, they can learn one protocol and use it every time.

This is where the “USB for AI” analogy fits. Without a universal standard, every device would need its own unique port and cable. Nightmare-inducing! In the same way, MCP give us a universal way for any LLM to talk with any tool.

Architecture

MCP has us sending POST requests over HTTP. Basically an MCP server is a tiny web service listening for JSON requests and responding with JSON results.

The Messages

That’s 80% of traffic in MCP. Here’s the tea.

getServerInfo: When a client connects to a server, it sends a handshake message to establish the connection and exchange capabilities.

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "getServerInfo"
}

Example response:

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "server_name": "NewsAPI",
    "version": "1.0.0",
    "capabilities": ["resources", "tools", "tools/list"]
  }
}

“capabilities” will always be an array with some combination of 1) “resources”, 2) “tools”, and 3) “tools/list”.

resources/list: What data do you have?

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "resources/list" # or "tools/list" or "events/list"
}

Example response:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": [
    {
      "uri": "news:headlines",
      "name": "Top Headlines",
      "description": "Latest top news headlines across all categories",
      "type": "feed",
      "format": "json"
    },
    {
      "uri": "news:category:sports",
      "name": "Sports News",
      "description": "Latest updates in sports",
      "type": "feed",
      "format": "json"
    },
    {
      "uri": "news:search",
      "name": "Search News",
      "description": "Search for news articles by keyword",
      "type": "search",
      "format": "json"
    }
  ]
}

The MCP spec requires certain fields, but doesn’t limit you to just those fields. You can add whatever fields you want.

resources/get: That thing you said you have? Give it to me.

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "resources/get",
  "params": {
    "uri": "news:headlines",
    "filters": { "category": "technology", "country": "tx" },
    "limit": 10
  }
}

Response:

{
  "jsonrpc": "2.0",
  "id": 3,
  "result": {
    "content": [
      {
        "title": "Rap Payne wins lottery again",
        "author": "Amanda Silberling",
        "source": "TechCrunch",
        "publishedAt": "2025-12-17T14:00:00Z",
        "url": "https://techcrunch.com/article123",
        "summary": "Rap Payne has once again defied the odds by winning the lottery. Can you believe the luck on this guy?"
      },
      {
        "title": "Breakthrough in Quantum Computing Announced",
        "author": "Megan Farokhmanesh",
        "source": "Wired",
        "publishedAt": "2025-12-17T12:45:00Z",
        "url": "https://wired.com/article789",
        "summary": "Researchers have made a significant advancement in quantum computing, potentially revolutionizing the field."
      },
      ... more articles ...
    ],
    "metadata": {
      "category": "technology",
      "country": "tx",
      "totalResults": 10,
      "fetchedAt": "2025-12-17T14:05:00Z"
    }
  }
}

That’s resources. Now for tools.

tools/list: What can you do?

{
  "jsonrpc": "2.0",
  "id": 4,
  "method": "tools/list"
}

Response:

{
  "jsonrpc": "2.0",
  "id": 4,
  "result": [
    {
      "name": "getWeather",
      "description": "Get current weather for a location",
      "inputSchema": {
        "type": "object",
        "properties": {
          "location": { "type": "string" }
        },
        "required": ["location"]
      }
    },
    {
      "name": "getStockPrice",
      "description": "Get latest stock price for a ticker",
      "inputSchema": {
        "type": "object",
        "properties": {
          "ticker": { "type": "string" }
        },
        "required": ["ticker"]
      }
    }
  ]
}

tools/call: Do this for me.

{
  "jsonrpc": "2.0",
  "id": 5,
  "method": "tools/call",
  "params": {
    "name": "getWeather",
    "arguments": {
      "location": "Arlen, TX"
    }
  }
}

Response:

{
  "jsonrpc": "2.0",
  "id": 5,
  "result": {
    "temperature": 92,
    "condition": "Partly Cloudy",
    "humidity": 60,
    "wind_mph": 8
  }
}

Final thoughts on MCP

This super quick overview and examples are enough to give you a good understanding of how MCP works. We made no attempt to cover every detail of MCP here. For all the official details, visit the official MCP site.

Thank God for MCP! It is going to make communication between LLMs and tools doable and enable agentic AI and more capable generative AI something that any of us can write.