In short. Model Context Protocol (MCP) is an open protocol for AI applications to access external data and callable tools through servers.

Definition: Model Context Protocol (MCP) is an open protocol that standardizes how AI applications access context and tools from external systems. An MCP server provides data or callable functions to a client within an AI application.

Key facts

The 2026-07-28 specification defines communication through JSON-RPC 2.0 messages, a structured format for requests and responses.
The MCP project's architecture overview distinguishes the host application, its clients, and the servers that supply context.
The resources specification identifies each resource by a URI, an identifier used to request its content.
The tools specification defines tools/list for listing available functions and tools/call for invoking one.

How does the exchange work?

The architecture overview describes a client-server arrangement. The host coordinates clients; each client communicates with a server. A server is a program that supplies context, whether it runs locally or remotely.

The communication mechanism is called a transport. Standard input/output, or stdio, carries messages between local processes on the same machine. Streamable HTTP carries client messages through HTTP POST requests and supports streaming through Server-Sent Events.

In this revision, requests are self-contained. Each carries protocol version and capability information in a _meta field, allowing the server to process it independently. The architecture documentation leaves decisions about how an application uses language models and manages the supplied context to the application.

What can a server expose?

The specification defines resources as contextual data, prompts as templated messages or workflows, and tools as executable functions. The architecture guide gives a database example: a server can offer a query tool, a resource containing the database schema, and a prompt with examples for using the tools.

The tools documentation gives a weather function as an example. This shortened definition preserves its name, description and input schema, which describes the accepted parameters:

{
  "name": "get_weather",
  "description": "Get current weather information for a location",
  "inputSchema": {
    "type": "object",
    "properties": {
      "location": {
        "type": "string",
        "description": "City name or zip code"
      }
    },
    "required": ["location"]
  }
}

Here, location is a required text parameter. A client lists tools to obtain their definitions, then invokes a selected function by name with its arguments. For a website operator considering an integration, these definitions make the implementation choices concrete: which data to expose and which operations to make callable.

What controls access to data and actions?

The specification's security principles require explicit user consent for data access and operations. Hosts must obtain consent before exposing user data to servers or invoking tools. The same document treats descriptions of tool behaviour as untrusted unless they come from a trusted server.

The resources specification requires servers to validate resource URIs and recommends access controls for sensitive resources. It also requires sanitizing file paths when serving filesystem resources to prevent directory traversal, where a path attempts to reach files outside its permitted location.

FAQ

Do you need MCP for a website to appear in Google's AI features?

As of October 2026, Google's AI features documentation states that AI Overviews and AI Mode have no additional technical requirements. Your page must be indexed and eligible to appear in Google Search with a snippet.

How can you read a resource exposed through MCP?

The resources specification defines resources/list to list available resources and resources/read to retrieve content using its URI. Your application determines how that content enters the model's context, including through explicit user selection or automatic inclusion.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer