Shopify just gave every store a free agent API. Here is what it still does not do

What actually went live

WebMCP lets a page declare what an AI agent is allowed to do on it, instead of leaving the agent to reverse-engineer the DOM. On a Shopify Liquid storefront, the declared tools cover most of a shopping session.

• Catalogue: search_catalog, browse_store, get_product, show_variant.
• Cart: get_cart, update_cart, cancel_cart.
• Checkout and orders: proceed_to_checkout, manage_orders.
• Content: search_shop_policies_and_faqs.

Two constraints keep this smaller than the headline suggests, and both are worth holding onto. The agent has to be running in the shopper's browser, in a Chromium build with the origin trial active, which today mostly means Gemini in Chrome. And the tools are registered at runtime through a JavaScript API rather than published at a discoverable address, so there is no directory of agent-ready stores and no way to audit a competitor's setup from outside.

The part that has not been solved

Picture the agent mid-session. It has called search_catalog and it has four waterproof jackets in front of it, three of them from stores that also got WebMCP for free on 21 August. It now has to pick one. Nothing in the WebMCP tool surface helps with that. The tools describe what a store can do, not why a product deserves the sale.

What an engine actually weighs when it chooses between comparable products is evidence: complete Product schema, an AggregateRating backed by real reviews, named and dated review text it can quote, and customer photos and video tied to the specific variant. None of that arrives with WebMCP. A store with immaculate agent tooling and four thin reviews is easier for an agent to transact with and no easier to recommend.

There is a second gap that gets missed. WebMCP only exists while an agent is on your page. A shopper asking ChatGPT for jacket recommendations without visiting any store is served from what the engine crawled and cached, and your in-browser tools are irrelevant to that conversation. That surface is reached by structured data, crawl access and a machine-readable index like llms.txt, which is the layer Idukki's agent feed publishes.

What to do about it this week

1. 1Confirm your storefront is Liquid. If it is, you already have the tools and there is nothing to install. If you are on Hydrogen, WebMCP is a developer preview rather than a default, so treat it as a build decision.
2. 2Stop budgeting for the read paths. Any 2026 roadmap line that reads "expose catalogue search to agents" is now partly funded by your platform. Move the sprint rather than spending it twice.
3. 3Audit the evidence instead. Product schema completeness, AggregateRating with a real count, named reviews with dates, and product-tagged UGC. Our free AI Visibility Check grades all of it, including agent tooling, on any URL.
4. 4Publish for the out-of-browser surface. An llms.txt index and an agent-readable feed reach the engines answering shoppers who never land on your site. That is the majority of AI-mediated demand today.
5. 5Read the implementation detail once. The sequencing across MCP, OpenAI tools and Google's agent feed is covered in the merchant's guide to MCP and agent tool calls, updated for this change.