Short answer
A store MCP server lets AI assistants search your products, check stock and look up orders through defined tools instead of guessing from your web pages. You may already have one: Shopify documents a UCP-based catalog and cart MCP endpoint for stores, and WooCommerce offers an MCP integration in developer preview for managing a store. Add one if you want agents to act on accurate data, and keep anything that moves money behind limits and review.
Key takeaways
- An MCP server gives AI assistants exact answers about your products and orders.
- Shopify and WooCommerce each offer an MCP endpoint, for different jobs.
- Keep order and payment tools behind limits, confirmation and logging.
What does an MCP server do for a store?
It gives AI assistants a set of tools for your store, such as "search products" or "get order status," so they can get exact answers instead of reading and guessing from your pages.
MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside tools and data. An MCP server is the piece a business runs. It tells a connected assistant what tools exist, what inputs each one needs, and returns structured results when the assistant calls them. If you are new to MCP, start with what MCP is.
For a store, typical tools are:
- search the catalog by keyword or category
- get one product's details, variants, price and stock
- build or update a cart
- check the status of an order or shipment
- answer policy questions, such as returns or delivery times
The value is accuracy. An assistant that reads a product page might miss that a size is out of stock. A tool that returns live stock does not.
Do Shopify and WooCommerce stores already have one?
Often yes, but they do different jobs, and both have changed during 2026.
Shopify. Shopify's developer documentation describes a storefront MCP endpoint for AI agents. Its original Storefront MCP, at /api/mcp on the store's domain, let agents search the catalog and manage carts. Shopify has since moved those tools to UCP-conforming versions at /api/ucp/mcp:
- Catalog tools there are
search_catalog,lookup_catalogandget_product. - On 24 June 2026 Shopify announced that the old cart tools were deprecated in favour of the new Cart MCP tools (
create_cart,get_cart,update_cartandcancel_cart), and said the old ones would be maintained until 31 August 2026. - Shopify's migration page now says the catalog and cart tools on the old endpoint have been removed and tells developers to "Use the Universal Commerce Protocol (UCP) instead."
WooCommerce. WooCommerce's MCP integration is a different kind of server. Its developer documentation says it is "currently in developer preview" and lets an AI client query, create, update and delete products, and query and update orders. That makes it a tool for the store owner's own assistant, not for shoppers. WooCommerce's own merchant guide, from August 2026, says its MCP helps AI assistants manage a store but does not yet handle shopper checkout. Setup uses the MCP Adapter plugin and a WordPress Application Password, and the documentation warns that order and customer operations "may expose personally identifiable information (PII)." For more on Woo, see how to let AI agents buy from your WooCommerce store.
Does your store need its own MCP server?
You probably need one if you want AI assistants to act on accurate, live data from your store, and your platform does not already give you that.
Ask yourself three questions:
- Does my platform already provide one? On Shopify, the UCP-based endpoint may already cover catalog and cart. Check what it exposes before building anything.
- What do I want agents to do? Answering product and order questions is low risk. Placing orders, issuing refunds or changing prices is not.
- Who will connect to it? Today, people usually add a remote server to their assistant by URL. Claude's documentation says a user or organisation owner can add one as a custom connector. OpenAI's documentation describes adding a custom MCP server in ChatGPT. Commerce standards can also point to it: UCP lists MCP as one of its transports, so a UCP integration can serve agents over MCP.
If your answer to the second question is "only answer questions," a read-only server is a small, safe step. If agents will place orders, plan the controls first.
What should a store MCP server include?
Clear tools, accurate data, and limits on anything that changes state. A practical checklist:
- Name and describe each tool plainly. The AI model reads the description to decide when to call a tool. "Search products by keyword, returns name, price, stock and URL" works better than "query."
- Return live stock and prices. A wrong answer from a tool is a promise you may have to honour.
- Separate read tools from write tools. OpenAI's guidance is that approval can be skipped for read-only tools but should stay on for tools that modify data.
- Never expose admin powers to shoppers. Product editing and customer data belong on an authenticated admin server, not a public one.
- Log every call. You want to know which agent asked for what, and what your server answered.
- Plan for the 2026 spec. The 28 July 2026 MCP revision removed sessions and the
initializehandshake and requires aserver/discovercall. Its list results also carry caching hints. If you build your own server, use a current SDK and check which revisions your target clients support.
What are the risks of running an MCP server?
The main risks are wrong answers, actions you did not intend, and data leaking to the wrong place.
- Prompt injection. OpenAI warns that content a server can access could contain hidden instructions that cause unintended actions. Product reviews and descriptions are text an attacker might plant instructions in.
- Over-broad tools. A tool that can cancel any order, rather than the caller's own order, is a problem waiting to happen.
- Personal data. Order and customer tools can return names and addresses. Keep them behind authentication and return only what each tool needs.
- Money movement. If a tool can place an order, it needs spending limits and a way to hold a purchase for a person to review. Our guide on letting AI agents buy from your Shopify store covers limits in more detail.
How Sill helps
Sill publishes the files AI agents read, including an MCP server for your site, an agent card, llms.txt and a catalog, and keeps them current. The MCP server supports the 2026-07-28 MCP spec and older AI clients at the same address, so you do not have to pick one. It shows you which AI agents visit and what they ask for and buy. When an agent places an order, your rules approve it, decline it or hold it for a person to review, and every decision is signed and recorded. Sill installs as a WordPress plugin, a Shopify app, on any site behind Cloudflare, or with one script tag. Run a free readiness scan to see whether agents can find an MCP server on your site today.
Questions people ask
Will an MCP server make my products show up in ChatGPT or Google?
Not on its own. Product listings in those surfaces come from their own programs, such as merchant feeds and Google Merchant Center. An MCP server helps an assistant that is already connected to your store get accurate answers.
Is my store's MCP server public?
It depends on the server. A shopper-facing catalog server is meant to be reachable by agents. An admin server, like WooCommerce's, requires credentials because it can read orders and customer details.
Do I need a developer to run an MCP server?
For a custom server, yes. If your platform provides one, or a service hosts one for you, setup can be a few settings. Testing it with a real AI client is still worth doing.
Sources
- Model Context Protocol: Architecture overview · checked 16 Nov 2026
- MCP specification 2026-07-28: Key changes · checked 16 Nov 2026
- Shopify developers: Migrate from Storefront MCP · checked 16 Nov 2026
- Shopify developer changelog: Storefront MCP cart tools are being deprecated in favour of UCP Cart MCP · checked 16 Nov 2026
- WooCommerce developer docs: MCP · checked 16 Nov 2026
- WooCommerce: From found to bought, getting your store ready to sell through AI · checked 16 Nov 2026
- UCP specification overview · checked 16 Nov 2026
- OpenAI developers: MCP · checked 16 Nov 2026
- Claude docs: Build an MCP server for Claude · checked 16 Nov 2026