Anthropic has released Claude for Commerce: open source reference code for a shopping agent and a merchant agent, released under Apache 2.0. It includes working examples across retail, travel, telecom and entertainment. A hard rule built into the merchant agent: no tool call moves money or changes the store without a human approving it first.
It's a capable release, and building a good agent on top of it is still real engineering work: prompting, evaluation, integration with your specific catalogue and systems, getting the guardrails right. That doesn't disappear because the reference code is free. What does change is the starting point. Work that used to take a platform team months to get to a working baseline is now available off the shelf.
Readiness matters more than which agent you run
A well-built agent still runs into the same wall once it reaches your site, and that wall has nothing to do with which agent it is.
Discoverability
Whether the agent can find the product at all. This means structured data and feeds an agent can query, not just pages built for a browser to render.
Understandability
Whether the agent can parse what it finds. Pricing locked in an image, specs split across tabs, a page that reads fine to a person and means nothing to a script.
Transactability
Whether the agent can finish the job. This is where most sites fail. A retailer can run a fluent, well-built AI shopping assistant with no working Cart API behind it, so the agent recommends a product, describes it accurately, and hits a wall the moment it tries to add it to a basket.
Where the numbers actually sit
Across the retailers we've scanned, the average Transactability score is 28.8 out of 100. Nearly a third score zero. Several of them already have some form of AI shopping assistant live today. It talks. It can't buy.
Snapshot from Aido’s scanned retailer cohort, not a market-wide benchmark. Scores describe the evidence available in each read-only scan and may change as coverage and methodology evolve.
The protocols check capability, not brand
UCP and ACP describe how an agent and merchant can exchange capability and checkout information. AP2 binds trust to a signed mandate from the user, while MPP operates at the payment layer. None of these declarations makes a request succeed by itself. What decides that is whether the merchant has a usable catalogue, cart API, checkout session, mandate handler, or another system capable of answering—and whether the relevant operation is actually tested.
Claude for Commerce makes it cheaper to build a good agent. It has no effect on whether your Cart API exists.
Building the agent got easier. What it has to work with didn't change.