Anthropic has released a commerce-agent blueprint for organisations that want to build shopping and merchant-facing assistants with Claude. Published on 2 September, the open repository is intended as a practical starting point rather than a new consumer shopping product. It packages reference implementations, integration patterns and operational guardrails for retail, travel, telecommunications and ticketing teams. Anthropic says the aim is to reduce the work needed to move an agent from a promising demonstration to a service that can work with real catalogues, policies and customer journeys.
The release is notable because it separates two roles that are often merged in demonstrations. A shopping agent sits in a company app or website and helps a customer find, compare and assemble products before handing the cart to checkout. A merchant agent works on the other side of the transaction, drawing on business data to answer questions about sales, inventory, pricing and promotions. The distinction matters: each role needs different data, permissions and safeguards, even when both are powered by the same model platform.
A reference build rather than a hosted storefront
The blueprint includes working examples for both roles. According to Anthropic, teams can build them with the Claude Messages API, Agent SDK or Claude Managed Agents, which remains in beta. The repository also includes a Claude Code plugin intended to help developers adapt the starting point to their own systems. Organisations retain responsibility for the underlying integrations, including catalogues, customer data, order history, checkout flows and the policies that determine what an agent may say or do.
That is an important boundary. The shopping example can connect a conversational interface to product search, multi-item planning, customer preferences and customer-service questions. It can present products, comparisons and a cart within the conversation, rather than returning a plain text answer. But it does not replace a merchant's payment process. Anthropic describes checkout as an integration point that stays with the retailer, whether the business uses its own process or an agentic-payments provider.
Controls are built into the proposed workflow
Anthropic presents the blueprint as a way to build useful automation without handing commercial decisions to an unrestricted model. The shopping-agent patterns are designed to ground prices and product information in catalogue data and to avoid manipulative upsell behaviour. The merchant-agent patterns cover sales analytics, inventory management, marketing and promotions, but the company says a person approves a suggested change before it goes live. That human approval step is a meaningful limitation, not a footnote: pricing, campaigns and inventory actions can all have immediate commercial consequences.
For teams evaluating agent frameworks, the release provides a concrete model of where tool boundaries should sit. The agent may retrieve products, assemble a suggested basket or flag a stock risk, while business systems remain the source of truth and a person retains authority over consequential actions. It also gives developers a place to review the assumptions made by the examples before connecting them to live customer data. Those controls still need testing against a company's own policies, edge cases, identity model and regional consumer rules.
Multiple deployment paths keep the blueprint portable
Anthropic says the code can be deployed through the Claude API, Amazon Bedrock, Microsoft Foundry or Google Cloud Vertex AI. That portability may matter to larger organisations that already standardise on a cloud provider or need to keep model access within an existing procurement and security arrangement. It does not make the design provider-neutral: the examples are built for Claude and rely on Anthropic's tools and development workflow. However, the ability to use supported cloud distribution channels gives buyers more flexibility over where inference is consumed and governed.
The company also positions the release alongside ecosystem partners including Accenture, Mastercard and Visa. These relationships signal that the blueprint is aimed at organisations that need implementation support and established payments or commerce infrastructure, rather than only small teams experimenting with a chatbot. Anthropic cites reported results from existing retail deployments, but those figures are company claims and should not be treated as a universal forecast. Outcomes will depend on product data quality, customer experience design, reliability, integrations and how carefully an organisation scopes autonomous behaviour.
What the release changes for commerce teams
The immediate change is access to an opinionated collection of code and patterns that can shorten the earliest stage of a commerce-agent project. Instead of beginning with a blank agent loop, teams can inspect examples for customer-facing search and support as well as internally facing merchant workflows. The accompanying live demonstrations and engineering material can help technical leaders judge the integration effort before committing to a production programme. The code is available publicly, so it can also be assessed by security, legal and product teams before customer data is connected.
It does not remove the hard work of deployment. Retail and travel systems contain changing availability, complex policies, customer-account data and payment boundaries; an answer that is plausible but wrong can affect trust quickly. Organisations will need to decide which sources the agent can use, how current they must be, where it should ask for confirmation and how staff can correct errors. They will also need observability for tool calls and outputs, particularly when the agent is allowed to build a cart or make a recommendation using commercial data.
Anthropic's blueprint is therefore best read as implementation guidance for a bounded class of agent applications, not an announcement that autonomous commerce is solved. Its strongest practical contribution is the explicit treatment of catalogue grounding, human approval and the separation between a customer assistant and an internal merchant assistant. For AI Provider Index readers, the release shows Claude expanding its developer-facing materials around agents that connect models to operational systems, with a clearer focus on safeguards at the point where conversation becomes a commercial action.