How it works
The protocol connects two sides:
- The MCP server, on the company's side, describes what it makes available: tools (search for a product, check stock), data (product pages, return terms) and prompt templates.
- The MCP client, built into the AI application, discovers those capabilities and calls them when the conversation needs them.
So the model no longer answers from memory: it queries the source and gets current data.
Before MCP, every connection between an AI assistant and a tool needed a custom integration. With a shared standard, one server can serve several compatible assistants, and one assistant can plug into many services.
Why it matters in e-commerce
AI shopping agents recommend what they can read. Without direct access, they rely on pages that may have been indexed weeks ago, with outdated prices or stock. An MCP server gives them structured, real-time access to the catalog, prices and availability.
For the brand, it's a way to stay in the picture when search, comparison and sometimes the purchase itself happen inside an AI assistant. The brand also controls what it exposes, and agent requests should be authenticated before they get anywhere near an order.
In its analysis of Claude for Commerce, iAdvize goes over the three ways Anthropic says agentic commerce is being built. In the first, the shopper talks inside an AI app like Claude or ChatGPT, and the retailer's systems get queried remotely during the conversation: that's what an MCP server is for. In the third, the brand builds its own agent on its own site. iAdvize shares Anthropic's reading: the site remains where the brand controls end to end what gets shown, recommended and sold. The two work together: the MCP server serves shoppers' agents, the AI Shopping Assistant serves visitors.
The same article lists three rules Anthropic sets for connecting an agent to a retailer's systems, worth applying before opening an MCP server:
- Shopper identity comes from the retailer's server, never from the browser or from the agent itself.
- The agent gets that shopper's permissions and nothing beyond them.
- Any content the brand didn't write itself (a customer review, for example) is information to read, never an order to carry out.
Real-life example
A shopper asks their AI assistant for "a carry-on under 16 inches that fits under the seat, delivered by Friday." The assistant queries a luggage brand's MCP server: it gets the exact dimensions, stock and delivery time, then shows two models at today's price. Without that server, it would have answered from indexed pages and might have suggested a model that's out of stock.
Common confusion
MCP doesn't replace APIs. An MCP server usually sits on top of a company's existing APIs and presents them in a form an AI model can discover and use on its own. Tech teams don't have to rebuild anything: they add a layer over what already exists.