What is an MCP (Model Context Protocol) server?

An MCP (Model Context Protocol) server is not a machine or a hosting plan: it is a role within the protocol. It is the program that offers capabilities to an AI assistant, wherever it runs: on your computer or in a vendor's cloud.⁠^{[1]Model Context Protocol, documentation. Architecture overview (version 2026-07-28)} If a platform announces that "your store now has an MCP server", or a vendor offers to "connect your catalog to Claude or ChatGPT", this is what they mean. A store should care because it is the door through which an AI agent queries its catalog or prepares a cart without opening the site, and because whoever controls that door decides what the agent can do inside.

How it works

The protocol's documentation defines servers as programs that expose specific capabilities to AI applications through standardized interfaces, and gives everyday examples: file, database, GitHub, Slack or calendar servers.⁠^{[2]Model Context Protocol, documentation. Understanding MCP servers (version 2026-07-28)} A server offers three kinds of piece, and each is controlled by someone different:⁠^{[2]Model Context Protocol, documentation. Understanding MCP servers (version 2026-07-28)}

• Tools, controlled by the model: functions it decides to call based on what the user asks, such as searching products or creating a cart.
• Resources, controlled by the application: read-only data that serves as context, such as a document or a database schema.
• Prompt templates (prompts), chosen by the user: ready-made flows for a task.

On the other side is the AI application (the host), which creates one client for each server it connects to and keeps a dedicated connection with it.⁠^{[1]Model Context Protocol, documentation. Architecture overview (version 2026-07-28)} Tools are what an integration uses most: each has a name, a description and a JSON schema with its parameters. The client asks for the list with tools/list and calls one with tools/call.⁠^{[3]Model Context Protocol. Tools (specification 2026-07-28)} And the server's responsibility is written into the specification: it must validate all inputs, control access, rate-limit calls and sanitize what it returns.⁠^{[3]Model Context Protocol. Tools (specification 2026-07-28)}

Local or remote

A server is called local when the application launches it on the same machine and they communicate over standard input and output (stdio); it usually serves a single client. It is called remote when it lives somewhere else and is reached over HTTP (Streamable HTTP); it usually serves many.⁠^{[1]Model Context Protocol, documentation. Architecture overview (version 2026-07-28)} The meaning of the messages is the same in both cases: the transport only decides how they travel.⁠^{[4]Model Context Protocol. Transports (specification 2026-07-28)} OpenAI, for the servers that give tools to a ChatGPT plugin, asks that they be deployed at a stable HTTPS address with Streamable HTTP, and protected with the specification's authorization if they touch private data or act on someone's behalf.⁠^{[5]OpenAI Developers. MCP server (ChatGPT plugins)}

Authorization is optional in the protocol, but it has rules: an HTTP server that uses it should follow the flow the specification defines, based on OAuth 2.1, and a local one should not follow it but take credentials from the environment.⁠^{[6]Model Context Protocol. Authorization (specification 2026-07-28)}

What changed for whoever operates one

The July 28, 2026 version of the specification mainly changed the life of the remote server. There is no longer an initial handshake or sessions: each request carries its protocol version and the client's details, so any instance behind a load balancer can serve it without shared storage. HTTP requests carry the method and the tool name in headers, so a gateway or a web application firewall can route and measure without reading the body, and tool lists say how long they can be cached.⁠^{[7]Model Context Protocol, official blog (July 28, 2026). The 2026-07-28 Specification}

In a store

A store has three ways of being behind an MCP server, and they are not mutually exclusive:

This is what a client would see when asking tools/list of a store's own server, a made-up example. The format is the specification's; the name, the description and the parameters are decided by whoever writes the server:⁠^{[3]Model Context Protocol. Tools (specification 2026-07-28)}

{
  "name": "search_products",
  "title": "Catalog search",
  "description": "Searches in-stock products by free text and maximum price",
  "inputSchema": {
    "type": "object",
    "properties": {
      "query": { "type": "string", "description": "What the customer is looking for" },
      "max_price": { "type": "number", "description": "Maximum price in dollars" }
    },
    "required": ["query"]
  }
}

The description is not decoration: it is what the model reads to decide when to call the tool, and OpenAI lists it among the fields that help it choose.⁠^{[5]OpenAI Developers. MCP server (ChatGPT plugins)} A vague description produces wrong calls, and a tool that returns outdated data produces outdated answers, however well built the protocol is.

How to check it

1. Review which servers each AI application your team uses has connected and whether each one is local or remote.⁠^{[1]Model Context Protocol, documentation. Architecture overview (version 2026-07-28)}
2. Look for the server in the official MCP registry. Names are tied to a verified GitHub account or domain, although the registry is in preview and leaves security analysis of the code to the package registries and the catalogs that consume it.⁠^{[10]Model Context Protocol, documentation. The MCP Registry}
3. Ask for the list of tools and read each description and each schema: what it can read, what it can write and with what parameters.⁠^{[3]Model Context Protocol. Tools (specification 2026-07-28)}
4. If the server is remote and touches store data, check that it requires authorization with the flow the specification defines.⁠^{[6]Model Context Protocol. Authorization (specification 2026-07-28)}
5. If you are the one operating it, test that it rejects out-of-schema inputs and that it rate-limits calls, as the specification asks.⁠^{[3]Model Context Protocol. Tools (specification 2026-07-28)}

Related concept: MCP

MCP is the protocol, the shared rules; the MCP server is one of its parts, the one that offers capabilities, just as the client is the one that uses them.⁠^{[1]Model Context Protocol, documentation. Architecture overview (version 2026-07-28)} The MCP entry covers the protocol and its history. Above the two are the agentic commerce protocols, such as UCP or ACP (Agentic Commerce Protocol), which set how an agent creates a cart and gets to payment. Shopify, for example, offers its cart and checkout for agents with UCP over MCP.⁠^{[8]Shopify, developer documentation. Migrate from Storefront MCP} And below is tool calling, the model's ability to ask for a function to be run.

To go further, ChatGPT Shopping shows how products reach an assistant today, and the agentic commerce reading path orders the rest of the protocols.