Claude Commerce Agents: an Apache-2.0 blueprint you own

Anthropic published Claude Commerce Agents, a reference blueprint for building shopping and merchant agents, under the Apache-2.0 license on GitHub. It is Python, it went up on September 1, and it is already past 1,600 stars. For anyone running a store, the useful part of this commerce agent blueprint is not the agent loop. It is the list of integration points it forces you to name.

What actually happened

The repo ships two reference implementations (Anthropic, github.com/anthropics/commerce-agents). A shopping agent handles natural-language product search, multi-item requests, personalized recommendations, cart building and in-conversation customer service, then hands the customer to checkout. A merchant agent analyzes sales, tracks inventory, recommends pricing and promotions and drafts marketing campaigns — with human approval required before anything goes live.

Both are built on the Messages API, the Agent SDK, or Claude Managed Agents (beta), and they run on the Claude API, Amazon Bedrock, Microsoft Foundry and Google Cloud Vertex AI. There is a Claude Code plugin for scaffolding an implementation against your own backend, plus self-guided demos across retail, travel, telecom and ticketing. Shopify published its own set of implementations on top of the blueprint, wiring a storefront agent over UCP and Sign in with Shop and a merchant agent over the Admin API.

Anthropic says enterprise customers running shopping agents have seen carts "up to 35% larger" and shoppers "60% more likely" to complete a purchase. Treat those as vendor-reported: no methodology, no baseline, no named account attached to either number. The code is verifiable. The lift is not.

Why an open commerce agent blueprint matters for your business

Read the integration list, because it is a bill of materials for your own data: catalog, cart, checkout, customer preferences, order history. Every one of those is a place where a small retailer's system is a spreadsheet, a plugin, or a field somebody types by hand. The blueprint assumes each is an API that answers correctly and fast. That gap — not the model — is why most agent pilots stall.

Apache-2.0 is the second thing worth noticing. The harness is yours to fork, and it deploys on four different clouds, which means the model is a swappable component and your product data is not. We build vendor-agnostic systems for exactly this reason: the labs will keep releasing better agents, and the only durable asset is a catalog clean enough that any of them can read it.

And keep the approval gate the merchant agent already has. An agent that recommends a price change and waits is a tool. An agent that applies one is an incident with a delay on it.

Sources: Anthropic, anthropics/commerce-agents on GitHub.