Anthropic has published the code for building Claude commerce agents, and released it under an Apache 2.0 licence that anyone can fork. The announcement went up on 2 September and the repository had drawn 2,000 stars and 339 forks by 5 September. There is no licence fee and no waiting list.
The pitch is about who owns the customer. Instead of sending your shoppers into somebody else’s assistant, you run the assistant yourself, on your own site, over your own catalogue. That is a different proposition from every agent story of the last year, and it is the reason this one is worth a retailer’s afternoon.
What Anthropic actually released
There are two agents in the blueprint, not one, and the difference matters when you scope the work. The shopping agent sits on the storefront: it searches the catalogue, compares options, answers questions and assembles a cart. The merchant agent sits behind it, on sales analytics, stock levels, product listings, pricing and promotions, with a human approval step in front of anything that changes a price.
The repository ships four working verticals built on fictional ACME data: retail, travel, telecommunications and ticketing. They run, which means a team can see the shape of the thing before committing to a build rather than reading a specification and guessing.
Deployment is flexible. The code runs against the Claude API, Amazon Bedrock, Microsoft Foundry or Google Cloud Vertex AI, and it ships with a Claude Code plugin. Requirements are modest on paper: Python 3.11 or newer, Node 22 and an API key, according to Anthropic.
The cart stops at your checkout
This is the part most of the coverage has got wrong, and it changes the compliance conversation entirely. The blueprint contains no payment protocol. The shopping agent builds the cart and hands the shopper to the store’s own checkout, where the payment happens exactly as it does today. There is no product catalogue in the box either, and no advertising layer.
For a retailer that is good news rather than a gap. Your payment stack, your card data handling and your reconciliation stay untouched, which means the project does not drag your finance and compliance teams into scope on day one. It also means the agent cannot make a sale on its own, so anyone modelling this as a fully autonomous storefront is modelling something that has not shipped.
Accenture, Visa and Mastercard are named in the announcement, and Shopify published its own code the same day for connecting Claude agents to Shopify stores. Worth reading precisely: Visa and Mastercard are taking the blueprint to their merchant networks, not embedding themselves inside it. Payments still run through the channels they always did.
The numbers Anthropic puts on it
Anthropic says carts assembled with an agent run up to 35 per cent larger, that shoppers are 60 per cent more likely to complete the purchase, and that small businesses selling through agents see 40 per cent more conversion. It also cites Accenture research finding that 85 per cent of consumers are already open to dealing with an AI agent.
Treat those as vendor figures, because that is what they are. Anthropic does not publish the sample, the sectors or the baseline behind them, and a conversion number without a control group is a claim rather than a result. They are useful for sizing an experiment and useless for signing off a business case.
The honest version for a board paper is narrower. A shopping assistant that knows your full catalogue removes the search friction that kills carts on sites with deep ranges. Whether that is worth 35 per cent on your store is something only your own test will tell you.

Why Australian retailers should look now
The Australian angle is not a case study, because no local retailer is named. It is timing. Shopify sits under a large share of small and mid sized Australian stores, and Shopify shipped its connector on day one, so the path from blueprint to a running pilot does not wait on anyone’s roadmap. Anthropic has had a Sydney office since earlier this year, which matters mainly for support and procurement conversations.
The catch is engineering, and Australian business press flagged it immediately: this needs a team that can already work with the Claude API or one of the big cloud platforms, as SmartCompany reported. Free to licence is not the same as free to run. The corner shop is not deploying this in a week.
For mid market operators the sensible first move is small. Point the shopping agent at one product category, measure cart size and completion against your current search, and keep the merchant agent out of pricing until you trust the approvals. That is the same discipline any customer facing AI build needs, and it is where most of the value is recovered.
A reference build, not a product
The most important sentence in the whole release is the one about support. This is a reference implementation. Anthropic is not maintaining it as a product, is not accepting outside contributions, and expects each merchant to fork it and own it from there. Whatever you deploy becomes your codebase to patch, secure and upgrade.
That is a real cost and it belongs in the estimate. An abandoned fork of a shopping agent is worse than no shopping agent, because it keeps answering customers with a catalogue it no longer understands. Budget for maintenance or do not start.
The wider shift is the one to watch. For two years the assumption was that agentic shopping meant customers living inside a handful of big assistants and retailers renting access to them. A free, self hosted blueprint points the other way, and the model choice underneath it starts to matter commercially rather than technically, which is the ground our comparison of Claude and ChatGPT for business workloads covers.
