Claude Commerce Agents: Anthropic's Official Blueprint for E-Commerce Agents

When you're building agent applications, what you lack most isn't models or frameworks — it's an industry reference you can copy and run.

How do big labs actually put agents into production? How should a shopping agent be designed? A merchant back-office agent? Papers and blog posts can't answer that as well as official, runnable code.

Anthropic shipped exactly that. Claude Commerce Agents (2980 stars, Apache-2.0, Python) is the official reference blueprint for building shopping and merchant agents with Claude.

What you get

Two finished agents, each owning one side:

• A shopping agent a business embeds in its app for customers — picking products, comparing, ordering;
• A merchant agent staff use to run the back office — products, inventory, orders, promotions.

And it's not a toy demo: it ships four runnable verticals (retail, commerce, telecom, entertainment), all built over the same libraries, so you can adapt one to your own business.

Three signals it's worth watching

• Safety is the first principle, not an afterthought. This is what I value most. The blueprint is explicit: checkout only renders the cart for the host to complete — it never actually charges; every merchant write is staged until a human approves it. Business rules, authorization, and compliance belong to the deployer. That "agents propose, humans approve the consequential actions" design is the only kind of agent an enterprise can actually trust.
• One definition, many runtimes. An agent is defined once (prompt, skills, tool contracts, gates) and runs on the Messages API, the Claude Agent SDK, and Managed Agents. That "define once, run anywhere" engineering is worth copying for anyone building an agent platform.
• It's official Anthropic. A reference implementation written by the company behind Claude — effectively the official best-practice answer, open-sourced. Far more trustworthy than secondhand community tutorials.

The honest caveat

It's a reference blueprint, not a turnkey commerce system. Every company, brand, product, and person inside is fictional (the only company is ACME); it doesn't place real orders, charge cards, or change live listings. Treat it as a textbook and scaffold for "how to land agents in commerce," then wire in your own real data and flows. It's also bound to the Claude ecosystem (Messages API, Claude Agent SDK), so weigh that against your current model stack.

I've localized the README and core docs to Chinese: https://github.com/yangshun2005/commerce-agents-cn

If you find this project useful, a star on the original repo supports the author's ongoing maintenance.