What is MCP (Model Context Protocol)?

MCP stands for Model Context Protocol, the protocol AI assistants use to connect to data and tools in other systems. If you have read that "your store needs an MCP server" or that an assistant "has connectors", this is what is being talked about. It matters to a store because it is one of the ways an assistant can query a catalog and create a cart without opening the website: Shopify, for example, already offers its catalog and cart tools for agents at an MCP address on every store.⁠^{[1]Shopify, developer documentation. Migrate from Storefront MCP}

How it works

The specification sums it up as an open protocol for integrating language model applications with external data sources and tools. It is inspired by the Language Server Protocol, which did the same for programming languages in code editors: a standard connection instead of a custom integration for every pair.⁠^{[2]Model Context Protocol. Specification (version 2026-07-28)} Messages are JSON-RPC 2.0, and there are three roles:⁠^{[2]Model Context Protocol. Specification (version 2026-07-28)}

• Host: the AI application that opens the connections, such as a desktop assistant.
• Client: the connector that lives inside the host, one per server.
• Server: the service that offers context and capabilities.

A server can offer three things: resources (data and context used by the user or the model), prompts (message templates and flows for the user) and tools (functions the model runs). The client, in turn, can offer the server elicitation: the server asks the user for more information in the middle of a task.⁠^{[2]Model Context Protocol. Specification (version 2026-07-28)}

Tools are the piece that touches a store most. Each one has a name, a description and a JSON schema with the parameters it accepts. The client asks for the list with tools/list and calls one with tools/call. They are designed to be controlled by the model, which decides when to use them based on the conversation, but the specification asks that there always be a person in the loop able to reject a call, and that the application show which tools are exposed and when they are invoked.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)}

Security is left to whoever implements it. The specification requires explicit user consent before exposing their data to a server or invoking any tool, treats tools as arbitrary code execution and acknowledges that the protocol cannot enforce these principles by itself.⁠^{[2]Model Context Protocol. Specification (version 2026-07-28)}

From Anthropic to an open foundation

Anthropic introduced MCP on November 25, 2024. The problem it pointed to was that every new data source required its own integration, and it released at the same time the specification, the development kits, support for local servers in Claude Desktop and a repository of ready-made servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer.⁠^{[4]Anthropic. Introducing the Model Context Protocol}

On December 9, 2025 it donated the protocol to a foundation for open agentic AI projects, the Agentic AI Foundation, a fund within the Linux Foundation created by Anthropic, Block and OpenAI. In that announcement it counted more than 10,000 active public MCP servers and adoption in ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code.⁠^{[5]Anthropic. Donating the Model Context Protocol and establishing the Agentic AI Foundation}

The current version was published on July 28, 2026 and changes the foundation: the protocol no longer has sessions, each request is complete and any server behind a load balancer can handle it, extensions become a formal framework, and Roots, Sampling and Logging are deprecated, with at least twelve months of support.⁠^{[6]Model Context Protocol, official blog. The 2026-07-28 Specification} If documentation talks about "initializing the session", it describes an earlier version.

In a store

The clearest case today is Shopify. Its old Storefront MCP server, at https://{store}/api/mcp, let an agent search the catalog and manage a cart. Those tools were removed from that address and are now offered with the UCP (Universal Commerce Protocol) at https://{store}/api/ucp/mcp, where each request has to include the agent's profile. The policies and frequently asked questions tool stays at the old address.⁠^{[1]Shopify, developer documentation. Migrate from Storefront MCP}

This is what a tool call looks like in a fictional example store with its own MCP server. The tool name and its parameters are defined by each server; the message format is the specification's:⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)}

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "search_products",
    "arguments": {
      "query": "6-cup induction moka pot",
      "max_price": 60
    }
  }
}

The response comes back with the result as text, as a JSON object or both, and if something fails, with an error the model can read to correct the call.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)} Clione has its own MCP server on the way, so an assistant can read the store's catalog and send answers and enrichments to Clione; it is not available yet.

How to check it

1. Find out whether your platform already exposes MCP. On Shopify, the tools for agents are at https://{store}/api/ucp/mcp.⁠^{[1]Shopify, developer documentation. Migrate from Storefront MCP}
2. Ask for the list of tools with an MCP client and read the name, description and schema of each one.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)}
3. Don't trust the descriptions of a server you don't know. The specification asks that a tool's annotations be treated as untrusted unless they come from a trusted server.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)}
4. Before connecting a server to an assistant, check what it can touch. If it can write to your catalog, the application has to ask your permission on every invocation.⁠^{[2]Model Context Protocol. Specification (version 2026-07-28)}

Related concept: tool calling

Tool calling is a model's ability to ask for a function to be run with certain arguments. MCP is the standard way of offering it those functions: where they are, how they are listed and how they are called, whatever the application.⁠^{[3]Model Context Protocol. Tools (Model Context Protocol, specification 2026-07-28)} A model can call tools without MCP, if the developer defines them by hand; MCP exists so the same tool works across many applications without being rebuilt. And above both sit the commerce protocols, which set which tools are needed to buy.

To keep going, the AI agent entry explains who uses these tools, agentic commerce explains what they buy for, and the agentic commerce reading path puts the rest of the protocols in order.