The assistant checks your stock by calling a tool you publish

You read that a live catalog hook-up can answer stock in chat, and you still want to know what that hook-up actually is. MCP (Model Context Protocol) is an open standard for connecting an AI application to outside data and tools during a conversation. Apps such as Claude or ChatGPT can connect that way when someone configures them to. The official overview is What is the Model Context Protocol?.

That is a different job from putting labeled facts on a public product page or in a feed. For when a page is enough and when it is not, see ChatGPT can look up your stock only after you connect it.

Three roles in one connection

The architecture guide names three participants:

The server is not JSON-LD on your webpage and not a row in Merchant Center. A stranger asking ChatGPT for a watch does not automatically reach your server until that app has been connected to it.

Servers can run on the same machine as the app (common for local tools) or on a remote address the client reaches over the network. The protocol describes both patterns; the details are in the architecture doc.

What the shop exposes for stock

For live price and availability, the piece that matters is usually a tool: a named function the model can invoke, with an input schema and a structured result. The tools specification defines how clients list tools (tools/list) and call them (tools/call).

Servers can also expose resources (readable data) and prompts (templates). A minimal catalog hook-up for stock often starts with one or two tools that read the same database or API your shop already uses for checkout.

Shape of a stock lookup (illustration)

Illustrative example, not a logged test. The Northloop Trail Watch is a fictional product at €300. This is not a chat transcript. It is the kind of labeled answer a get_product tool might return when the assistant passes a SKU (the shop’s own product code) your server understands:

{
        "sku": "NL-TRAIL-S-BLK",
        "name": "Northloop Trail Watch S",
        "price": "300.00",
        "priceCurrency": "EUR",
        "availability": "InStock"
      }
      

Your real server would define the tool name, required arguments (for example sku), and the exact fields in the tools schema. The assistant only sees what that call returns in this session, after the connection exists.

What “create it” means in practice

You do not install MCP on a product page. You build and run an MCP server that wraps catalog access you already trust.

1. Define tools that read live price and stock from the same source as your shop floor, such as the shop API, the database, or the PIM, and keep them aligned with your webpage, JSON-LD, and feed.
2. Run the server where the client can reach it: on a developer machine for experiments, or on a hosted URL for a remote client.
3. Connect the client in the host app. The official Build a server walkthrough uses a small example server and Claude Desktop as the host. That guide is the right place for SDK steps and config files; this article stays at the shop-owner map.

Support bots, internal catalog assistants, and dev tools such as Cursor are the realistic first hosts. Do not assume every public ChatGPT shopping session will call your tools without that wiring.

What this still will not do

Publishing an MCP server does not make a third-party assistant recommend you to everyone who asks a generic buying question. That limit is the whole point of ChatGPT can look up your stock only after you connect it.

MCP also is not robots.txt, not llms.txt, and not a product feed by itself. It is in-session access for an app that someone already connected to you.

Read next if: Checkout in chat comes after the shortlist · Your product lives in three places