What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open standard that lets AI applications such as Claude, ChatGPT, or coding environments access external data sources, tools, and services in a uniform way. Anthropic released MCP in November 2024, and since December 2025 the protocol has been housed at the Agentic AI Foundation under the umbrella of the Linux Foundation. An MCP server describes once what a system can do, and every MCP-capable AI application can then use those capabilities.

Term profile

How does the Model Context Protocol work?

The Model Context Protocol connects three roles: host, client, and server. The host is the AI application a person works with, such as a chat window or a coding environment. For each connected server, the host starts its own client, which maintains exactly one connection. The server is a program that provides a large language model with the data and functions of a specific system, for example a database, a calendar, or an SEO tool.

A typical request runs in 4 steps:

1. Discover: The client asks the server with tools/list which tools it offers. Each tool comes with a name, a description, and a JSON schema for its inputs.
2. Decide: The language model reads the descriptions and decides, based on the prompt, whether it needs a tool and which one.
3. Call: The client sends tools/call with the matching arguments to the server.
4. Respond: The server returns the result, and the model works it into its answer.

Local servers run on the same machine and communicate via standard input and output (stdio). Remote servers run at the provider and are reachable via Streamable HTTP, and the specification recommends OAuth for authentication.

What building blocks does MCP define?

MCP defines 3 building blocks a server can offer: tools, resources, and prompts.

• Tools: Executable functions the AI application can call, such as a database query, a search, or creating a calendar entry.
• Resources: Data sources that provide context, such as file contents, records, or responses from an interface.
• Prompts: Reusable templates that structure an interaction with the model, such as a workflow for a monthly report.

In the other direction, a server can ask the user for input through elicitation, for example to request a missing detail or a confirmation before an action. With the 2026-07-28 specification, MCP has also become stateless: Every request carries its protocol version and capabilities itself. The earlier sampling feature, which let a server request text responses from the client’s model, has been marked as deprecated since then.

What is the difference between MCP and an API?

An API is the interface of a single service, while MCP is a shared format in which many such interfaces are described for AI applications. An MCP server usually sits in front of an existing API and translates its functions into tools that a language model finds and understands at runtime.

MCP is also distinct from retrieval augmented generation. RAG pulls matching passages from a knowledge base into the context. MCP can do that too through resources, and it additionally allows actions in other systems.

What role does MCP play in SEO and AEO?

For SEO and answer engine optimization (AEO, also called GEO), MCP matters in two directions: As a working tool for analysis and as an access point through which AI agents act on websites and stores.

As a working tool, MCP connects an AI assistant directly to SEO databases, Search Console, ad accounts, or the content management system. Rankings, search volumes, and campaign data can then be queried and combined in conversation. Several major SEO tools offer their own MCP servers for this.

As an access point for agents, MCP changes how customers buy and book. In agentic commerce, AI agents search, compare, and buy on a person’s behalf, and store platforms provide their own MCP servers for it. Since February 2026, Google Chrome has been testing WebMCP, a draft from the W3C Web Machine Learning Community Group that lets a web page register its functions directly in the browser as tools for agents. How this shifts search engine optimization is covered in the glossary entry on agentic SEO.

For content, MCP changes little at the foundation. An AI crawler still reads public pages as HTML, and an llms.txt summarizes a website as a static file for language models. An MCP server adds an interface to both that delivers data on request and performs actions. If you want to be mentioned in AI answers, you therefore first need citable content and a clear entity, and for stores, bookings, and software you also need a way for agents to take action. Both are part of the work on visibility in AI systems.

How does an SEO agency use MCP day to day?

In its daily work, taismo connects several MCP servers to its AI assistant, including an SEO database, the Google Ads account, a website builder, and a social media planner. This work has produced 3 rules:

1. Keep permissions as narrow as possible: Ad accounts are connected read-only. A model that can read budgets has no need for access to change them.
2. Double-check the numbers: An MCP server passes on the values of the service, including faulty ones. A keyword difficulty of 0 combined with a high cost per click, for example, is a missing value. Before a decision, the value is therefore checked against the actual search results page.
3. Plan for outages: Every server is an additional dependency. If it does not respond, the assistant lacks the tool, and the query runs through the direct route to the interface.

How AI agents are changing research, analysis, and execution in SEO is shown in the magazine article What AI agents really do for your rankings.

What are the security risks of MCP?

The biggest security risk with MCP is that a language model takes instructions from tool descriptions and tool results just as seriously as the user’s instructions. This leads to 3 known attack paths:

• Prompt injection: A document, web page, or email returned by a server contains hidden instructions to the model.
• Tool poisoning: A manipulated server hides instructions in the description of a tool. Security researchers publicly described this pattern in April 2025.
• Vulnerabilities in server software: MCP servers are programs with their own weaknesses. In July 2025, CVE-2025-6514 disclosed a critical vulnerability in the widely used helper program mcp-remote.

Four measures reduce the risk considerably: Install servers only from trusted sources, give each server only the permissions it needs, have a person confirm write actions, and update servers regularly.

Frequently asked questions about the Model Context Protocol

What is an MCP server?
An MCP server is a program that provides the data and functions of a system in the format of the Model Context Protocol. It runs locally on your own computer or as a remote service at the provider.

Is the Model Context Protocol free?
Yes. The protocol and the official SDKs are open source. Costs arise from the connected services, such as query quotas of an SEO database or the use of the language model.

Does my website need its own MCP server?
For a purely informational website, a dedicated MCP server is usually not necessary, because well-readable HTML and structured data carry its visibility. For stores, booking systems, and software where AI agents are supposed to take actions, it is worth looking into.

Which AI applications support MCP?
MCP is supported by Claude, ChatGPT, Visual Studio Code, and Cursor, among others. OpenAI adopted the protocol in March 2025, and Google DeepMind announced support in April 2025.