Claude commerce agents are getting an open development blueprint from Anthropic, giving retailers and other consumer businesses reference software for shopping assistance and store operations. Released September 2, the package includes two agent types and examples for retail, travel, telecom, and entertainment.

 

The launch gives engineering teams a starting point for connecting conversational AI to their own services ahead of holiday shopping. Its practical boundary is equally significant: the reference agents prepare transactions and proposed changes, while the business retains responsibility for completing purchases and approving operational actions.

 

Claude Commerce Agents Serve Shoppers and Store Teams

In its launch announcement, Anthropic describes a shopper-facing assistant that can work inside an existing website or app. It can help people find products, compare options, assemble a cart, and ask questions about orders or store policies without leaving the conversation.

 

A second agent serves merchant staff. Its tasks include examining sales, highlighting inventory issues, suggesting promotions, and drafting campaigns. Anthropic also names Shopify, Priceline, Visa, Mastercard, and Accenture among customers or ecosystem participants discussing the wider commerce effort. Their roles differ; the announcement does not describe a single shared deployment.

 

That separation puts two different business problems under one release: helping a customer decide what to buy and helping an operator decide what to change. A successful shopping answer and a sound inventory recommendation require different information, even when both use the same underlying model.

 

Reference Code Keeps Payments Outside the Demo

The public repository offers implementations for the Messages API, Claude Agent SDK, and Claude Managed Agents. A commerce-builder plugin for Claude Code can help scaffold a store-specific agent, add flows, author evaluations, or review an existing implementation.

 

The reference package makes several boundaries explicit:

  • Demo companies, people, products, and brands are fictional.
  • Examples do not place real orders or charge payment cards.
  • Checkout presents a cart for the host application to complete.
  • Merchant writes remain staged until a person approves them.

 

The repository also says its example applications have no authentication. Production teams therefore need to connect identity, authorization, and their own business rules before exposing an agent to customers. The code is an implementation reference, not evidence that a particular retailer has passed a security review.

 

For deployment teams, that distinction changes the work estimate. Running a demonstration can establish that an interface functions; integrating a real catalog and testing account permissions establishes whether the same interface can operate responsibly inside a business. Those are separate milestones.

 

Further Reading

 

Anthropic Puts Enforcement in Application Code

Anthropic’s accompanying engineering guide recommends keeping most commerce conversations in one agent loop, with reusable skills supplying specialized procedures. The stated aim is to preserve shared shopping context and avoid unnecessary handoffs between separate agents.

 

More consequentially, the guide places enforcement outside model instructions. Its design checks whether product or listing identifiers came from server responses, caps permitted changes, and requires approval before staged merchant actions take effect. It also rechecks limits when a change is applied.

 

The guide advises evaluating completed tasks rather than judging an agent only by a convincing conversation. Testing should include contradictory histories, missing data, and situations where an assistant must refuse or request clarification. Speed and model cost are measured alongside whether the requested outcome was achieved.

 

This approach reflects a basic commercial problem: a fluent explanation cannot establish that a price, refund, or available quantity is correct. The application must supply authoritative records, and the final action must remain consistent with them. A persuasive answer attached to the wrong item is still a failed transaction.

 

Retailers Retain the Storefront and Deployment Work

On its commerce product page, Anthropic describes the release as an open reference implementation rather than a supported product. Businesses fork and own their versions; the company says the reference code has no service-level agreement and is not maintained or open to contributions.

 

The same page positions Claude as an intelligence layer within merchants’ existing relationships. Anthropic says the offering does not take ownership of catalogs, supply chains, or delivery, and does not include advertising or paid placement in agents built from the blueprint.

 

There is a broader distribution contest behind the release. Digital Commerce 360 reported that Claude became the leading AI referral source for 15 of North America’s Top 1000 online retailers in the second quarter, up from one in the first. ChatGPT remained the dominant source in that dataset.

 

Referral traffic and an assistant embedded in a merchant’s own app measure different things. The former describes how shoppers arrive; the latter concerns what happens after they get there. Retailers assessing this blueprint will need their own evidence on completed purchases, accurate support, and operating costs before treating a working demo as a commercial result.