MCP Concepts

Every AI app re-invents tool connections. MCP — the Model Context Protocol — ends that: one open standard to plug models into tools and data. USB-C for AI.

▶ Watch this reel

What you'll learn

  1. What & why
  2. Host, client, server
  3. Primitives: tools, resources, prompts
  4. Transports & vs function calling

Remember this

What & why

Roles

Primitives

Transports & comparison

Code: Capability discovery: what a server announces

// What a server exposes at connect time (simplified):
{
  "tools": [
    { "name": "search_tickets",
      "description": "Search support tickets by keyword. Use for customer-history questions; NOT for live account data (use get_account).",
      "inputSchema": { "type": "object",
        "properties": { "query": {"type": "string"}, "limit": {"type": "integer", "default": 5} },
        "required": ["query"] } }
  ],
  "resources": [
    { "uri": "ticket://recent",  "name": "Recent tickets",
      "description": "Last 50 tickets as markdown" },
    { "uri": "kb://article/41",  "name": "Refund policy" }
  ],
  "prompts": [
    { "name": "ticket_summary",
      "description": "Summarize a ticket thread with sentiment" }
  ]
}

// The host turns tools[] into the model's function-calling schema —
// AG-02's description discipline, now riding a standard wire format.