Anthropic's fake store demos how Claude commerce agents should work

What it does

Anthropic open-sourced a two-sided blueprint for AI commerce agents built on Claude. One agent faces customers to search, compare, and fill carts; the other faces staff to analyze performance, adjust listings, and draft campaigns. Four runnable verticals—retail, travel, telecom, and entertainment—demonstrate both roles over the same libraries, though every company and transaction is strictly fictional.

The interesting bit

The merchant agent cannot actually change anything live. Every write is staged until a human approves it, and the checkout flow merely renders a cart for the host system to complete. The README is admirably blunt that this is reference code, not a product, and explicitly states it is unmaintained and accepts no contributions.

Key highlights

• Three runtime paths: raw Messages API loop, Agent SDK wrapper, or Anthropic Managed Agents via MCP manifest
• Shared commerce-common library handles fencing, memory, grounding, and safety rules across both shopping and merchant roles
• Shopping skills cover search, comparison, planning, cart, and memory; merchant skills cover analytics, listings, inventory, pricing, and campaigns
• Backend interfaces are designed for you to wire to your own catalog, cart, order, and analytics systems
• Safety docs map each enforced rule to its module, though the examples ship without authentication

Caveats

• Explicitly unmaintained and does not accept contributions
• No official MCP connectors ship; integration with warehouses, finance, or delivery platforms is left to your backend implementation
• The examples bind MCP servers to loopback and lack authentication, so they are demos, not deployment templates

Verdict

Worth studying if you are designing conversational commerce interfaces and need a sane starting point for tool contracts and safety gates. Skip it if you are looking for a drop-in Shopify plugin; this is architecture, not glue.