Anthropic open-sources reference blueprint for commerce AI agents

Anthropic just dropped commerce-agents on GitHub, a runnable reference blueprint for building full-cycle e-commerce AI agents with Claude. Instead of giving you another generic chat template, it delivers two dedicated agents: a customer-facing Shopping Agent embedded in storefronts and a back-office Merchant Agent for operators.

Anthropic notes that retailers testing Claude shopping agents saw cart sizes jump up to 35%, while consumer checkout completion rates rose by 60%.

What it can do

The Shopping Agent acts as an attentive digital salesperson. It can search and compare items, plan complex purchases, explain store policies, and load up the cart while remembering customer preferences across turns.

The Merchant Agent sits in the back office. It reviews performance trends, catches inventory alerts, drafts promotions, and modifies listings. The repository ships with 10 production-style skills split evenly across both agents.

To prove this works across different business models, Anthropic bundled four runnable vertical apps: general retail, travel with date-bound bookings, telecom with plan matrices and fee disclosures, and entertainment ticketing with timed holds and seat maps. Each vertical contains both a customer storefront and a merchant portal.

Safety gates and running it locally

The smartest architectural choice here is how Anthropic handles financial risk. Neither agent is allowed to execute real-world transactions autonomously.

Nothing places an order, charges a card, or alters live catalog listings. At checkout, the agent merely compiles the cart and passes a handoff URL back to your existing checkout pipeline. On the merchant side, every write action is staged; prices or stock levels will not change until a human clicks approve or types y in the console.

You will need Python 3.11+ and Node 22 to get started. The project includes seven local Python packages, eight web apps, and three runtime targets: the Messages API, the Claude Agent SDK, and hosted Managed Agents over MCP. Claude Code users can also install the commerce-builder plugin to scaffold a new storefront assistant with a single slash command.

The entire blueprint is released under the Apache 2.0 license on GitHub. Keep in mind that Anthropic explicitly labels this as an unmaintained reference project that will not accept outside pull requests.

Why it matters

E-commerce teams often get stuck when moving AI from marketing copy to actual operations. Either the assistant is just a shallow FAQ bot, or the business fear of unauthorized discounts and hallucinated stock updates keeps it off production servers. This release shows how to safely sandbox Claude around real inventory and pricing data.

For platform developers and retail engineers, this repository is well worth cloning today just to inspect the tool contracts and approval gates. Still, we think you should treat it as an architectural guide rather than turnkey software. The sample databases are populated with fictional ACME goods, so the heavy lifting will still be implementing custom backend interfaces for your own catalog, payments, and warehouse APIs.