Anthropic Open-Sourced the Agent That Shops Your Store
Anthropic just did something unusual: it open-sourced the shopping agent that will buy from your store. Claude Commerce Agents is an Apache-2.0 blueprint - a working shopping agent and a merchant agent, with the prompts, skills and tool contracts to run them. For a merchant that is a rare gift. You do not have to guess what an AI agent wants from your storefront anymore. You can read the code.
“We’re going to begin to use these agentic applications as these kinds of personal shoppers ... Agentic personal shoppers are predicted to be the future of shopping in some ways — able to discover, buy, and compare products for consumers more effectively.”
Most agentic-commerce coverage is speculation about what agents will want. This is the actual reference agent, published by the company that makes Claude and MCP, under a licence that lets you open it up. Read what its shopping agent reads, and you have your requirements list straight from the source instead of a vendor's guess.
What Anthropic actually shipped
Anthropic open-sourced Claude Commerce Agents (the repo is anthropics/commerce-agents), and it is two agents. The shopping agent searches, compares, plans a multi-item cart, checks out, answers order and policy questions, and remembers what a customer told it. The merchant agent explains performance, fixes listings, acts on inventory alerts, sets prices and promotions, and drafts campaigns. There are runnable reference builds for retail, travel, telecom and ticketing, and the same code deploys on the Claude API, Amazon Bedrock, Microsoft Foundry or Google Vertex. It is a blueprint, not a product, which is exactly why it is useful to read.
“The truly disruptive impact of agentic AI will hit retail supply chains and operations first, such as autonomous agents handling real‑time inventory rebalancing, dynamic pricing and vendor negotiations at scale, because that’s where the ROI is measurable,” said Walton.
Anthropic names Visa, Mastercard, Accenture and Shopify as launch partners, and each sits where you would expect. Visa's Intelligent Commerce and Mastercard's Agent Toolkit already open their networks to Claude, the payment-trust layer an agent needs before it can pay, and Accenture runs a multi-year enterprise-AI partnership with Anthropic that puts Claude in front of large clients. The buyer, the rails and the integrators are lining up together.
What the shopping agent reads from your store
Here is the part worth your time. The shopping agent runs against a StorefrontBackend, an interface a deployment implements over four systems: catalog, cart, order and policy. In plain terms, the agent expects to read your product catalog, operate a cart, look up orders, and find your terms, shipping and return policies. Miss one or leave it messy and the agent has less to act on. That is the same lesson as the feed and policy gates in the platform requirements piece, now confirmed by the reference agent itself rather than argued from the outside.
On Shopify, the agent shops through UCP and MCP
Shopify published matching examples (Shopify/claude-for-commerce-examples), and they are specific about the plumbing. The storefront shopping agent shops through Shopify's UCP endpoints - the Universal Commerce Protocol Shopify launched with Google - over an MCP client, with Sign in with Shop for personalized results. The concrete requirement is blunt: the store must serve a domain at /.well-known/ucp. The merchant agent works over the Admin GraphQL API with scoped tokens. That /.well-known/ucp path is not a footnote; it is the door the agent knocks on, and if nothing answers, there is no shopping to do.
This is exactly what an agent-readiness score measures
If you have followed our agent-readiness audits, this closes the loop. The checker scores MCP and UCP because those are the signals an agent uses to connect, and Anthropic's blueprint just made that literal: its shopping agent connects through UCP and MCP or it does not connect at all. The DTC brands we scored at 3 of 9 were passing MCP and UCP by inheriting them from Shopify, which is the reason this agent could shop them in the first place, and the reason an off-platform store scoring 0 could not. See the beauty batch for what that spread looks like across ten real stores.
The merchant agent runs on governed autonomy
One design choice is worth copying wholesale: the merchant agent never writes directly. Every change - a price, a listing, a restock - is staged in a ledger and applied only after a human approves it. That is the governed-autonomy pattern, and Anthropic building it into the reference implementation is a strong signal for how much leash to give an agent that can touch your store. If you let an agent run operations, stage and approve; do not hand it write access and hope.
What to do before Claude's agent knocks
- Expose the protocols. On Shopify, confirm your store serves /.well-known/ucp and the storefront MCP - the door this agent uses. Off Shopify, you self-host both.
- Clean the four systems. Catalog, cart, order and policy are what the shopping agent reads. Complete product data, a working cart, order lookup, and clear shipping and return policies.
- Check your own score. Our agent-readiness checker probes the same MCP and UCP signals the blueprint connects through, so run it and see what the agent would find on you.
- Read the code. It is Apache-2.0. Point an engineer at anthropics/commerce-agents and the Shopify examples to see the exact tool contracts, rather than waiting for a summary.
Agentic commerce stopped being a forecast the moment the reference agent went open source. The company behind the protocol on your storefront published the buyer and told you what it reads. The only question left is whether your store answers when it knocks, which is the whole point of the pillar guide.
Run the agent-readiness checker and see whether Claude's shopping agent could shop your store today.
FAQ
What are Claude Commerce Agents?
Claude Commerce Agents is an open-source (Apache-2.0) blueprint Anthropic published for building commerce agents with Claude. It includes a shopping agent (search, compare, plan a cart, check out, answer order and policy questions) and a merchant agent (performance, listings, inventory, pricing, campaigns), plus prompts, skills and tool contracts, with reference builds for retail, travel, telecom and ticketing. The repo is anthropics/commerce-agents.
What does the Claude shopping agent need from my store?
The shopping agent runs against a StorefrontBackend that a deployment implements over four systems: catalog, cart, order and policy. It expects to read your product catalog, operate a cart, look up orders, and find your terms, shipping and return policies. On Shopify it connects through the store's UCP endpoints and an MCP client, so the store must serve a domain at /.well-known/ucp. Incomplete or missing data in any of those systems gives the agent less to work with.
How does the Claude shopping agent connect to a Shopify store?
Per Shopify's own reference examples, the storefront shopping agent shops through Shopify's UCP (Universal Commerce Protocol) endpoints over an MCP client, with Sign in with Shop for personalization. The concrete requirement is that the store serves a domain at /.well-known/ucp. The merchant agent works separately, over the Admin GraphQL API with scoped tokens, and stages every write for human approval.
Is Claude Commerce Agents free to use?
The blueprint is open source under Apache 2.0, so the code is free to read, run and adapt. Running it calls the Claude API and needs an ANTHROPIC_API_KEY (API usage is billed), and it can also deploy on Amazon Bedrock, Microsoft Foundry or Google Vertex. It runs locally on Python 3.11+ and Node 22. It is a reference blueprint to build from, not a finished hosted product.
How do I know if my store is ready for the Claude shopping agent?
The agent connects through MCP and UCP, the same protocols an agent-readiness check scores. Run your store through GEOlikeaPro's agent-readiness checker to see whether those signals are present and whether AI crawlers can even reach you. On Shopify most of that is inherited from the platform; off-platform you have to self-host it. Then confirm your catalog, cart, order and policy data are complete, since those are the four systems the shopping agent reads.