Anthropic opened the plans for its commerce agents. Checkout stays with the merchant.
Claude, Shopify, Visa, Mastercard and Accenture share the launch announcement. They do not share the same job. This is a precise map of what shipped, what still needs integration and what the early numbers cannot yet establish.
On 2 September, Anthropic did not put a new autonomous buyer on the market. It published the reference code and architecture rules a merchant can use to build one.
That sounds like a narrow distinction, yet it moves the entire responsibility boundary. Claude can interpret a request, search a catalogue, compare products and prepare a cart. Product, price, stock, customer identity and payment still come from merchant systems.
Two agents, on opposite sides of the counter
The shopping agent
It searches, compares, assembles multi-item requests, remembers preferences and prepares the cart. It can also answer order, return and refund questions from connected policies.
Boundary: it hands the cart to checkout. It does not charge a card.
The merchant agent
It explains sales, flags inventory risk, prepares listing edits and proposes prices, promotions or campaigns from store data.
Boundary: every write stays staged until a person approves it.
What the repository ships, and what it deliberately leaves out
| Layer | Shipped by Anthropic | Left to the deployment |
|---|---|---|
| Reasoning | Agent loop, prompts, skills and tool contracts | Model choice, latency and cost budgets |
| Data | Backend interfaces | Real catalogue, price, stock, orders and policies |
| Actions | ID provenance, caps and staged changes | Permissions, human approval and business rules |
| Payment | Cart handoff to checkout | Authentication, authorization, charge and recourse |
The code runs through the Messages API, Claude Agent SDK or Claude Managed Agents, which is still labelled beta. Examples cover retail, travel, telecom and ticketing or entertainment. A Claude Code plugin can scaffold an agent, add a flow, author evals or review an existing agent.
The project uses the Apache 2.0 licence. Its README also says this is an unmaintained reference implementation that does not accept contributions. It is not a managed service or a support promise.
The system around the chatbot is the real release
In the demo, the model never receives a general right to write. The server records the identifiers it returned during the session. A cart accepts only a product from that set. A merchant change can be applied only through a server-generated ID approved in a real operator surface.
Limits are checked again when the change is applied. A discount that was valid when Claude proposed it can be refused minutes later if stock, budget or policy has changed. For regulated disclosures, the server supplies approved copy. The model selects it but does not rewrite it.
“The model’s most dangerous action is to propose.”
Anthropic, architecture guide, 2 September 2026
Five companies, five different roles
Logos are used only to identify the companies. Names open their official websites; X accounts were checked against company websites or public newsrooms on 6 September 2026. Vector marks: Simple Icons 16.30.0.
Anthropic calls Accenture, Mastercard and Visa solutions and ecosystem partners that can help clients and merchant communities use the blueprint. The source does not call them co-authors of the code.
“Trust must remain at the center of every transaction.”
Jack Forestell, Visa, quoted by Anthropic
Shopify and Priceline: what is established, and what is not
Anthropic says Shopify and Priceline run customer-facing agents built with Claude. Priceline says the latest generation of Penny uses Claude to reason across the many parts of a trip. This confirms Claude is used in production commerce agents. It does not prove those companies deployed the new repository unchanged.
Shopify provides more specific evidence. Its public repository connects a storefront agent to UCP and Sign in with Shop, and a merchant agent to the Admin GraphQL API. Merchant operations retain their approval gate.
What this changes for an online seller
The first consequence is organizational: catalogue, stock, prices and policies are no longer only displayed on a page. They become the context an agent reasons over. Anthropic therefore recommends connecting agent tools to existing business systems instead of rebuilding their logic in the model. For a merchant, answer quality then depends directly on the freshness of those sources and the precision of the fields the server returns.
The second concerns the customer relationship. Visa describes an approach in which the purchase finishes on the merchant storefront, preserving customer data, transaction visibility and the brand relationship. An agent can therefore move discovery and comparison upstream without immediately making the site or checkout obsolete. This is a documented consequence of Visa’s integration model, not proof that every journey will evolve this way.
Sources for this analysis: Anthropic · Visa Intelligent Commerce Connect
The blueprint, plugin and implementations are public.
Anthropic reports carts up to 35% larger and shoppers 60% more likely to complete a purchase for some customers.
The announcement publishes no sample, timeframe or comparison method for those two figures.
The right first experiment is not an autonomous purchase
The blueprint reduces the time needed to test a useful interface. It does not lower the evidence bar for changing a price, stock level or payment. A prudent path starts read-only, then adds carts and staged changes, while keeping the existing checkout until authorization and recourse work end to end.
That is less spectacular than a bot buying on its own. For a merchant, it is also the part that can actually be evaluated.