Anthropic's Commerce Agent Blueprint: What Merchants Need Before a Claude Shopping Agent Goes Live

What Anthropic Actually Released

On September 2, 2026, Anthropic published a blueprint for building commerce agents on Claude: an open GitHub repository with working reference agents, plus a Claude Code plugin for adapting them to a specific catalog and set of policies. It is not a marketplace and not a new checkout. Anthropic is not selling products or acting as a merchant. It is handing retailers and platforms a starting point for agents they run themselves.

The blueprint ships two agents:

  • A shopping agent that lives inside your own app or website. Anthropic lists its capabilities as catalog search, multi-item planning, deep research, personalization, customer care, and in-conversation UI such as product cards.
  • A merchant agent for your internal team, covering sales analytics, catalog and inventory management, marketing and promotions, and dashboards inside an admin portal.

Both can be built on the Claude Messages API, the Agent SDK, or Claude Managed Agents (in beta), and deployed through the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. Anthropic named Accenture, Mastercard, and Visa as partners, and listed companies including Shopify, Priceline, Square, Klaviyo, and Wix as building with Claude. Commerce platform commercetools has already published a tutorial for wiring the blueprint to its backend.

This matters even if you never deploy one yourself. Platforms and agencies are building on the same patterns, so the requirements below are the ones your store will meet one way or another.

The Most Important Design Choice: The Agent Only Says What Your Catalog Says

Anthropic's guardrails constrain prices and products to actual catalog data. The shopping agent is not supposed to invent a product, guess a price, or "remember" a spec from training data. It also avoids manipulative upsell patterns, and suggestions from the merchant agent need human approval before they go live.

That is the right call for trust, and it changes where the risk sits. When an assistant answers from the open web, a weak product page means you get described badly. When an agent answers only from your catalog, a weak record means the agent has nothing to say at all. A missing attribute is not described wrongly. It just never comes up.

So the practical question for any merchant is not "is Claude good at shopping?" It is "what will my catalog give the agent when a shopper asks a specific question?"

What the Agent Needs From Your Backend

The blueprint lists its integration points as catalog, cart, checkout, customer preferences, and order history. The commercetools tutorial makes the catalog requirements concrete, and the lessons apply to Shopify, BigCommerce, Adobe Commerce, and WooCommerce just as well.

1. Prices that resolve for the shopper's market

In the commercetools example, price selection depends on currency and country settings. If they are not configured, products come back with no price. Any multi-currency or multi-market store has some version of this risk. An agent that cannot resolve a price cannot recommend the product with confidence.

2. Names and copy in every locale you sell in

The same tutorial warns that asking for a single locale returns nothing when that exact key is missing, so the backend has to fall back gracefully. If you sell internationally, gaps in translated titles become gaps in what the agent can show. Our guide to international AI search optimization covers how to audit them.

3. Searchable attributes that match how shoppers talk

Filters on attributes that are not marked searchable silently return zero matches. If a shopper asks for "waterproof, under 300 grams" and those attributes are not indexed, the agent finds nothing, and the shopper concludes you do not carry it. Map the attributes customers actually name in questions, reviews, and support tickets, and make sure they are filterable, not just written in description prose.

4. Variants that are real choices

The tutorial advises that variants should represent real choices and that identical variants be presented as one product. Duplicate or meaningless variants clutter the agent's options and make comparisons noisy. See product variant data for AI search for a cleanup approach.

5. Availability that means what it says

In commercetools, a missing availability value means no inventory record exists, which is not the same as out of stock. Every platform has edge cases like this: preorder items, made-to-order products, items stocked only at certain locations. If your inventory signals are ambiguous, an agent can either hide in-stock products or promise ones you cannot ship.

6. Discounts and tax setup that do not break carts

When product discounts are active, the agent has to quote the discounted price. And in the commercetools setup, every buyable product needs a tax category, or cart creation fails. These are mundane configuration details, and they are exactly where an agent conversation breaks at the moment of purchase.

7. Policies the agent can quote

Answering policy questions is a core job for the shopping agent. Shipping times, return windows, warranty terms, and exclusions need to exist as clear, current text the agent can retrieve, not scattered across banners and PDFs.

Checkout Stays Yours

The blueprint's shopping agent builds a cart and hands it to checkout. It does not take payment. Anthropic leaves that to the implementer: your existing checkout, or an agentic payments provider. That keeps you as merchant of record and keeps your fraud, tax, and fulfillment logic where it already lives.

It also means checkout friction still counts. If the cart the agent built arrives at a checkout with surprise shipping costs, a forced account creation step, or a broken discount, the conversation's goodwill is lost at the last step. Test the handoff the way a shopper would experience it, on mobile and as a guest.

How This Fits With ChatGPT, Gemini, and Shopify Agents

The blueprint is about agents on your own properties. It sits alongside, not in place of, the external surfaces covered elsewhere on this blog: ChatGPT checkout through the Agentic Commerce Protocol, Google's UCP-powered checkout in AI Mode and Gemini, and the Storefront MCP endpoint every Shopify store already exposes.

The common thread is that all of them read structured product data instead of interpreting your page design. One clean catalog serves every one of these surfaces. Separate fixes for each do not scale.

A Pre-Launch Checklist

Before you, your platform, or an agency puts a Claude-based shopping agent in front of customers:

  1. Pull a sample of real shopper questions from search logs, support tickets, and reviews.
  2. Run each through your catalog API and check whether the answer is actually in the data: price, availability, the specific attribute asked about.
  3. Mark the attributes shoppers filter on as searchable, and normalize their values (one unit system, consistent naming).
  4. Collapse duplicate variants and fill missing variant-level data.
  5. Verify pricing and naming for every market and locale you sell in.
  6. Confirm every buyable product can be added to a cart without configuration errors.
  7. Publish policies as structured, current text.
  8. Walk the cart-to-checkout handoff as a guest on mobile.
  9. Keep a human approval step on anything the merchant agent proposes, as the blueprint intends.

Where to Start

If you are not sure how your catalog holds up when an agent queries it directly, start with an AI visibility audit to find the gaps, then work through product optimization to fill the attributes, variants, and policy data agents depend on. The agent framework is now open source. The catalog behind it is the part only you can fix.

Want to see how AI engines perceive your brand?

Get an Audit for $87