There are two ways an AI agent can reach your store, and most of the conversation only covers one of them.
The server-side path is UCP: an agent connects to an MCP endpoint, searches a catalog, builds a cart. The other path runs entirely inside the browser, on the page the buyer is already looking at. It is called WebMCP, and on Shopify it is switched on by default.
Key takeaways
- WebMCP is a web standard for pages to register structured tools with an in-browser agent. Interface:
document.modelContext, withgetTools()andexecuteTool(). - Shopify registers them automatically on every Liquid storefront and on Hydrogen 2026-10 storefronts that render
ShopifyScripts. - Roughly 19 tools across five groups — catalog, cart, checkout, navigation and store info.
- Official framing: agents receive data instead of parsing screenshots.
search_shop_policies_and_faqsis the one live outlet for your knowledge base today.
The tool list
Group | Tools | What the agent can do |
|---|---|---|
Catalog |
| Search within the page, browse collections, switch variants |
Cart |
| Operate the live session cart directly |
Checkout |
| Advance to checkout and finish |
Navigation and orders |
| Move around the store and manage orders |
Store info |
| Answer policy and FAQ questions on the page |
Why replacing screenshots with function calls matters
For the past two years, browser agents operated ecommerce sites visually: screenshot, identify the button, compute coordinates, click, screenshot again to confirm. That loop is slow, expensive and brittle. A redesign breaks it. A cookie banner breaks it. A lazy-loaded price breaks it.
WebMCP replaces it with a function call, which changes three things for merchants.
Stability. The agent depends on a tool contract, not on pixel positions. Redesign your product page and the purchase path keeps working. For teams who ship frequently, this quietly removes a whole category of invisible breakage.
Cost. One add_to_cart call replaces a dozen screenshots and clicks. Cheaper and faster for the agent means agents complete more purchases rather than abandoning midway.
Observability. Tool calls are, in principle, measurable in a way simulated clicks never were. For the first time there is a path to knowing how many agents did what in your store — though as we note in the attribution gap, the reporting side of this is still immature.
The clause that makes it matter
Your store's own /agents.md contains a line that is easy to skim past: "Do not fall back to scripted form fills or page controls."
When tools are unavailable, agents are instructed not to resume clicking — they should hand the buyer back to a browser instead. Combine that with a storefront registering zero WebMCP tools and the outcome is specific: the agent does not degrade to the old visual method. It leaves.
Which turns the grep result into a real number. curl -s --compressed https://yourstore.com/ | grep -ci 'modelContext\|webmcp' returning zero is not a missing enhancement; it is a closed door. On the headless store we probed, that count was zero while a comparable Liquid store returned five matches — the full comparison is here.
The knowledge base connection
search_shop_policies_and_faqs deserves separate attention because of what it consumes: the facts you publish in Sales channels → Agentic → Knowledge base.
The protocol-side Ask capability, ask_business, is still a UCP draft and was absent from the 15 tools a live store returned to us. But this browser-side tool already reads the same facts. A buyer browsing your store with an in-browser agent can get your approved answers today.
That is the practical argument for building the knowledge base now rather than waiting for the protocol to stabilise. The asset is already in use.
What to actually do
- Verify registration. Run the grep. Greater than zero, or nothing in this article applies to you.
- If you are on Hydrogen, confirm you are on 2026-10 and that
ShopifyScriptsis rendered, orinitializeShopifyScriptsis called. This is the single most common cause of a zero count. - Test the tools that matter. Browse your own store with an in-browser agent and try: find a product, add it to the cart, ask a returns question. Three minutes, and it tells you more than any audit report.
- Publish knowledge base facts, since
search_shop_policies_and_faqsis live and reading them. - Do not assume default equals working. Heavily customised themes, unusual checkout modifications and third-party cart apps can all interfere.
FAQ
Is WebMCP a Shopify invention? No. It is a developing web standard for pages to expose structured tools to in-browser agents. Shopify's contribution is shipping an implementation across its storefronts by default, which is what turns a standard into a channel.
Does WebMCP replace UCP? No, they are parallel. UCP is server-to-server and powers cross-store discovery and agent-side checkout. WebMCP works on the page a buyer is already on. A full picture needs both — see the six-layer stack.
Can I register custom tools? Not through the default Shopify registration, which provides a fixed set. Treat the standard tools as the contract and invest in making them return good data.