Quick answer
Anthropic has released an Apache-2.0 reference implementation for building customer-facing shopping agents and staff-facing merchant agents with Claude. The repository includes shared tools and safety gates, four runnable industry examples, three runtime paths, and a Claude Code plugin for scaffolding or reviewing an implementation. It is code that a business must fork, connect to its own systems, secure, test, and operate—not a hosted storefront, checkout service, or supported product. The design deliberately keeps payment and business-changing actions outside direct model control.
The blueprint includes two agents and four runnable verticals
Anthropic's repository contains a shopping agent for customers and a merchant agent for staff. The shopping agent can search and compare catalog items, plan multi-item purchases, build a cart, answer order and policy questions, and remember customer-provided preferences. The merchant agent can analyze sales, surface inventory issues, recommend pricing or promotions, maintain listings, and draft campaigns. Runnable examples cover retail, travel, telecom, and entertainment, with storefront and merchant-portal surfaces for each vertical.
Teams can use the Messages API, Agent SDK, or Managed Agents
The same prompt, skills, tool contracts, and gates can run through a custom loop on the Messages API, through the Claude Agent SDK, or as a hosted Claude Managed Agent. Anthropic says the code can be deployed through the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. A bundled Claude Code plugin can scaffold a new commerce agent, add a flow, author evaluations, or review an existing agent. These are implementation paths, not a claim that every runtime has identical hosting, security, or support terms.
Anthropic recommends one agent loop with tools and modular skills
The accompanying engineering guide uses one model in a standard agent loop instead of placing an intent router and domain-specific subagents around every task. Frequently needed instructions, including critical safety, legal, brand, and user facts, stay in the system prompt; less common procedures load as skills. Anthropic reports that this pattern performed better than one large prompt or routine subagent handoffs in its enterprise comparisons, but it does not publish enough methodology for that result to be treated as an independent benchmark. Narrow research tasks and established specialist agents remain stated exceptions.
Sources: Anthropic
Typed tool calls carry commerce UI and business-system access
The design calls existing search, ranking, cart, inventory, pricing, analytics, and campaign systems through tools rather than asking the model to recreate their business logic. Product carousels, itineraries, plans, charts, and other interface elements are also represented as typed tool calls: the server validates and enriches the call, then the client renders it. That leaves a structured record in conversation history, although teams must account for buffering, schema retries, streaming behavior, and the latency of every connected service.
Sources: Anthropic
The model proposes actions, while the harness controls irreversible steps
In the reference implementation, no model tool call directly places an order, charges a card, issues a refund, changes a price, or launches a campaign. Checkout renders a cart for the host application to complete. Merchant writes create staged changes with server-generated IDs and require approval through an operator-facing surface before they can be applied. Limits are checked again at apply time, and write or render tools accept only IDs previously issued to that session. Deployers still own authentication, authorization, compliance, business rules, and production hardening; the example applications explicitly do not provide production authentication.
Long-term memory belongs in the application's data layer
Anthropic's guide treats cross-session memory as application data rather than model state. It recommends storing small, typed facts in a queryable database, enforcing allowed categories on the write path, separating users even when merchant accounts are shared, and providing ways to view, correct, and delete saved information. Retention periods and a regional or deployment-level off switch are also part of the suggested design. Those controls are especially important because preferences, account details, and inferred shopping context can be personal data.
Sources: Anthropic
This is a starting point, not a maintained commerce product
Anthropic describes the repository as an open reference implementation under the Apache License 2.0. Businesses fork it and maintain what they build; Anthropic says the reference code is not maintained, does not accept contributions, and has no service-level agreement. A realistic pilot should begin with read-only catalog or analytics flows, connect server-side backends with session-scoped credentials, add authentication, define approval and payment boundaries, test prompt-injection defenses, run task-level evaluations, and review privacy obligations before enabling memory or writes.
The blueprint is not a model or feature inside Chat AI
This release is developer reference code for teams building on Anthropic's platform and supported cloud routes. It does not announce a new Claude model or a prebuilt commerce-agent feature in Chat AI. Chat AI users can still use available models for tasks such as comparing requirements, drafting evaluation cases, or reviewing product copy, but connecting a live catalog, cart, payment flow, or merchant system requires a separate application and its own controls.
Frequently asked questions
What readers usually ask
What did Anthropic release for commerce agents?
Anthropic released an open reference implementation containing shopping and merchant agents, shared libraries, safety gates, four industry demos, deployment examples, tests, and a Claude Code plugin.
Can the reference shopping agent charge a customer?
No. In the reference design, the agent builds and displays a cart, but the host application owns checkout and payment. The backend interface exposed to the agent has no charge method.
Can the merchant agent change prices or launch campaigns by itself?
Not in the reference implementation. It stages proposed changes, and an approved server-side path must apply them after a person or policy authorizes the specific server-generated ID.
Which industries have runnable examples?
The repository includes retail, travel, telecom, and entertainment examples, each with a customer-facing storefront and a merchant-facing portal.
Is Anthropic's commerce-agent code a supported product?
No. Anthropic describes it as an Apache-2.0 reference implementation that teams fork and maintain themselves, without a service-level agreement for the reference code.
Is the Claude commerce-agent blueprint available in Chat AI?
It is not a Chat AI model or built-in feature. It is source code for developers to integrate with their own catalog, checkout, analytics, inventory, and policy systems.
Evidence