MCP Server Cards: How AI Agents Discover Your Site's Tools

The web has layers of discovery standards built up over two decades — robots.txt, sitemaps, Open Graph, JSON-LD. Now a new layer is forming for a new kind of visitor: AI agents that don't just read content but take actions. The MCP Server Card is how those agents discover what your site can do.

What Is the Model Context Protocol?

The Model Context Protocol (MCP) is an open standard developed by Anthropic that defines how AI models connect to external tools, data sources, and services. Think of it as USB-C for AI: a single standard interface that lets any agent connect to any tool without custom integration code for each pair. An AI agent supporting MCP can call search tools, read databases, execute code, post to APIs — anything — as long as the tool exposes an MCP-compliant server.

What Is an MCP Server Card?

Before an agent can use your MCP server, it needs to discover it and understand what it offers. The MCP Server Card is a JSON document published at /.well-known/mcp.json that answers these questions: What is this server? What can it do? How does an agent authenticate? What endpoint should it connect to?

Following RFC 8615, the .well-known directory is the standard location for machine-readable metadata files. By placing the server card here, any agent that knows the domain can discover MCP capabilities with a single predictable request.

Structure of an MCP Server Card

A minimal but complete server card looks like this:

{
  "name": "Acme Docs MCP",
  "description": "Access Acme product documentation, search articles, and fetch page content.",
  "version": "1.0.0",
  "endpoint": "https://acme.com/mcp",
  "protocol": "mcp",
  "protocolVersion": "2024-11-05",
  "auth": {
    "type": "oauth2",
    "authorizationUrl": "https://acme.com/oauth/authorize",
    "tokenUrl": "https://acme.com/oauth/token",
    "scopes": ["docs:read"]
  },
  "tools": [
    {
      "name": "search_docs",
      "description": "Search the Acme documentation by keyword",
      "inputSchema": {
        "type": "object",
        "properties": {
          "query": { "type": "string", "description": "Search query" }
        },
        "required": ["query"]
      }
    },
    {
      "name": "get_page",
      "description": "Fetch the full content of a documentation page by URL",
      "inputSchema": {
        "type": "object",
        "properties": {
          "url": { "type": "string", "description": "Page URL" }
        },
        "required": ["url"]
      }
    }
  ]
}

Key Fields Explained

• endpoint — The URL where the live MCP server accepts connections. Agents send JSON-RPC requests here.
• protocolVersion — The MCP spec version your server implements. Agents use this to determine compatibility.
• auth — How the agent should authenticate. Can be none, apikey, or oauth2. For OAuth2, links to your token endpoint.
• tools — An array of tool definitions with name, description, and JSON Schema input. This is what agents read to know what they can do.
• resources (optional) — Data sources the agent can read, separate from callable tools.

How Agents Use the Server Card

The discovery flow is straightforward: an agent that knows about your domain fetches /.well-known/mcp.json, reads the tool definitions, authenticates if required, and connects to the endpoint. From that point it can call any declared tool. No custom plugin code, no manual API key configuration — the entire handshake is automated from the server card.

This matters for search and citation too. AI-powered search engines like Perplexity increasingly understand MCP, and a published server card signals that your site is designed for agentic interaction — not just passive reading.

What Tools Should You Expose?

• search — Let agents query your content by keyword or topic
• get_page / fetch_content — Return the full text of a specific page
• list_products / list_articles — Let agents enumerate your catalogue or content inventory
• get_pricing / get_availability — Structured data tools for e-commerce use cases
• create_ticket / submit_form — Action tools for service-desk or lead-gen workflows

Announcing Your MCP Server via Link Header

Publishing the file at /.well-known/mcp.json is the primary discovery method, but you can also advertise it via an HTTP Link header for agents that check headers before probing .well-known:

Link: </.well-known/mcp.json>; rel="mcp-server-card"

Adding this header to your site's responses means any agent that inspects the HTTP headers — even before navigating to any page — can discover that an MCP server exists.