Shopify Catalog MCP: global discovery, storefront scope, and safe agent design

Shopify Catalog MCP is a machine-facing discovery interface, not a magic storefront. It gives an agent structured access to product discovery capabilities. The surrounding system still needs identity, policy, intent handling, confirmation, and transaction state.

Choose the correct catalog scope

Shopify documents two interfaces in About Catalogs:

• Global Catalog searches across Shopify merchants.
• Storefront Catalog searches within a single merchant store.

Use global scope when the product is a multi-merchant discovery experience. Use storefront scope when the user has already chosen a merchant, is browsing one brand, or the business needs a controlled assistant on its own surface.

Scope is a product decision, not only an endpoint choice. A global agent must explain merchant selection and comparable attributes. A storefront agent must avoid implying it searched the whole market.

Treat MCP as capability discovery

The MCP interface lets an agent discover and call structured tools. The Shopify agentic commerce documentation describes a broader journey: negotiate identity, discover products, build carts, create checkout, and monitor orders.

That means product discovery is one state in a longer transaction. Keep these states explicit:

1. user intent and constraints;
2. product candidates and source scope;
3. selected variant and live availability;
4. cart state;
5. buyer confirmation;
6. checkout handoff or completion;
7. order state and exceptions.

An agent should not reuse an old product result as if price and availability were still current at checkout.

Add trust boundaries before convenience

Host and reference the required agent profile. Request only the capabilities needed. Log tool calls without exposing payment credentials or personal data. Make buyer confirmation visible before a consequential transaction.

The public UCP specification includes identity and payment-related capabilities. Those are reasons to narrow permissions and validate signatures, not reasons to let one model prompt directly control every step.

Validate product answers

Build a representative query set from real catalogue decisions:

• use case and recipient;
• size, fit, material, or compatibility;
• budget and currency;
• delivery destination and deadline;
• stock and variant constraints;
• return or subscription conditions.

For each query, record retrieved products, missing relevant products, incorrect attributes, stale availability, and unsupported claims. Compare the answer with the visible page and the current admin record.

Design failure paths

Every tool call needs a defined response for timeout, empty results, conflicting data, capability mismatch, and partial checkout state. The safe fallback is usually to explain the limit, preserve user context, and hand off to the merchant or a conventional checkout. It is not to invent a product fact so the conversation can continue.

For KumoCart, this topic belongs in the strategic frontier: it matters before broad search volume appears, but future articles must stay tied to current documentation and reproducible agent tests.