
What Is Claude Commerce?
Carlos GarciaPublished What Is Claude Commerce?
If you searched for Claude Commerce expecting a shopping feature inside Claude, the first thing worth knowing is that it is not one. There is no button in the Claude app that buys things for you, and Anthropic does not describe one anywhere on its own pages.
What exists instead is something more interesting and much more specific: a set of open, published blueprints that let a business build its own shopping and merchant agents on top of Claude, and run them inside its own app. Anthropic wrote the reference code, put it on GitHub under an open license, and then said in plain terms that it is not a product and that it is not maintained.
That gap between what people expect and what was actually shipped is where most of the confusion lives. This guide covers what Claude Commerce is, what the two agents do, the one question everybody asks about payment, the limits Anthropic states outright, and the things Anthropic conspicuously does not say.
What Is Claude Commerce, Exactly?
Claude Commerce is an open reference implementation of two commerce agents built on Claude, published by Anthropic on September 2, 2026. Anthropic's own commerce solutions page answers the category question without hedging: "It's an open reference implementation, not a supported product."
The launch article frames it the same way. "Today, we're launching a blueprint to help build commerce agents on Claude," it says, and the solutions page carries the heading "An open blueprint for commerce agents on Claude."
The blueprint contains two agents, described in the repository as "Two commerce agents built on Claude: a shopping agent a business embeds in its app for customers" and "a merchant agent its staff use to run the back office." One faces your shoppers. The other faces your own staff.
Worth noting before you go searching: the name itself is unstable across Anthropic's own properties. The GitHub repository is titled "Claude Commerce Agents." The solutions page headline reads "Claude for commerce." The launch article is titled "Building commerce agents with Claude." All three are the same thing, and none of them is the canonical name, which is why search results for this topic look so scattered.
Is your own site ready for the traffic an AI agent sends you? Get a free SEO audit and find out what agents and crawlers actually see.
Free audit
Where do you stand in AI search?
A manual SEO + AI search audit of your site: how ChatGPT, Gemini, Perplexity and Claude read and cite it today, and what to fix first. Report within 48 hours, no credit card.
Get my free auditOr have it all done: AI SEO services, $999 once
Why Did Anthropic Publish a Blueprint Rather Than a Product?
The short answer is that commerce is where a model's judgment stops being the hard part. A shopping conversation is easy to demo and brutally hard to operate, because the moment an agent can do anything consequential, somebody has to own the authorization, the business rules and the liability. Anthropic's framing puts that ownership squarely on the deploying business.
The audience follows from that. The solutions page describes the goal as helping companies "build and deploy consumer and merchant AI agents in your app," and the deep-dive article is explicit about who it is written for: "This post is for the engineers and engineering leaders building these (or other consumer facing) agents."
So Claude Commerce is not sold to shoppers. It is written for the engineering team at a retailer, travel company, telecom or entertainment platform, and shoppers only ever meet it as a feature inside that company's own app.
What Does the Shopping Agent Do?
The shopping agent lives in your product, not in Claude. Anthropic's article says it directly: "The shopping agent lives inside your app or website."
Its documented jobs are discovery, cart assembly and customer service in one conversation. On search, the blueprint describes the ability to "Search the catalog and assemble the right set of items, including multi-item requests." On the cart, it is blunter still: "Build the cart and hand it to checkout."
It also absorbs the support conversation that normally happens somewhere else entirely. The agent can "Answer customer service questions in the same conversation, like where an order is, how to return or exchange an item." Note the verb: it answers questions about returns. No Anthropic page says the agent processes a return or issues a refund itself.
What Does the Merchant Agent Do?
The merchant agent is the back-office half, used by a business's own staff rather than by customers, and its default posture is the most quietly important design decision in the whole blueprint.
It does not act on its own. "By default, it drafts and stages changes, then waits for your review," the documentation says, and the solutions FAQ repeats the guarantee from the other side: "Nothing reaches your storefront until a person approves it."
If you have ever been nervous about pointing a language model at a live product catalog, that is the sentence to read twice. The staging-plus-approval pattern is not a setting someone remembered to switch on; it is the shipped default.
Does Claude Handle the Payment?
No. This is the most consequential fact about Claude Commerce and the one almost every secondary article gets fuzzy about, so it is worth quoting Anthropic several times over.
The launch article says the blueprint "leaves payment to you, whether that is your existing checkout or an agentic payments provider." The architecture deep-dive is stronger: "No model tool call moves money or changes the business."
Stronger again, and this is the line that actually explains the design: "the backend interface the agent calls has no charge method at all." The capability is not withheld from the model by instruction or by prompt. It does not exist in the interface the model can reach.
The repository summarizes the same boundary for all three risky verbs at once: "Nothing places an order, charges a card, or changes a live listing." And the solutions FAQ states the positioning in one sentence: "Claude is your intelligence layer, not a storefront or checkout."
Mechanically, the hand-off works by URL. The backend returns a checkout link and the host application renders it, and the model never sees that link. In the reference implementation the checkout tool simply "renders the cart with a button to place the order."
A human, in other words, still presses the button. If you are evaluating agentic commerce for risk, fraud or compliance reasons, that is the architectural answer, and it comes from Anthropic's own repository rather than from anyone's interpretation of it.
Most ecommerce analytics problems are measurement problems long before they are AI problems. Request a free audit and start with the data.
How Do You Actually Build On It?
You run it on Claude through one of three surfaces. Anthropic names them as the "Messages API, Agent SDK, or Claude Managed Agents (beta)," and the article notes the reference code also runs against the Claude API through Amazon Bedrock, Microsoft Foundry and Google Cloud Vertex AI.
The code itself is on GitHub as anthropics/commerce-agents, published under the Apache License 2.0. That license matters more than the star count: Apache 2.0 means you can take it, fork it, modify it and ship it commercially without asking.
The integration point is deliberately indirect, and this is where people misread the architecture. The agents do not call your systems directly. In the repository's words, "both agents reach your systems through the backend interfaces," and where a commerce platform has its own connector, "A commerce platform's own MCP server for catalog, cart, or checkout is called from a backend method server-side."
Read that carefully, because it inverts the usual mental model of agent tooling. The model is not handed a set of live commerce tools. Your backend holds the tools, and your backend decides what the model is allowed to ask for.
Which means the real work of adopting Claude Commerce is not prompt engineering. It is writing the backend interface that sits between the agent and your catalog, your cart and your order system, with your own authorization inside it.
Anthropic also names who to call if you would rather not build it alone, describing "our solutions and ecosystem partners such as Accenture, Mastercard, and Visa" as "working with us to enable clients and merchant communities to leverage the blueprints." Separately, its pages carry customer and partner quotes from Shopify, Priceline, Klaviyo, Square, Wix, Fetch, Intuit and Zomato.
Be precise about what those names mean, though. A quoted executive is not an integration. Anthropic does not state what the Visa or Mastercard collaboration actually includes, and no Anthropic page says Claude processes payments through a payment network.
When Should You Use It, and When Should You Not?
Use it when you already intend to build an agent into your own storefront and you want a credible starting architecture rather than a blank repository. The separation of model, backend and checkout is the genuinely valuable part, and it reflects design decisions that are expensive to discover yourself.
Use it as an architecture reference even if you never ship a line of it. The no-charge-method rule, the staged-changes default and the server-side connector pattern are reusable conclusions, and they apply to any agent you point at a transactional system.
Use the merchant agent pattern if your risk conversation is stalling. "Drafts and stages changes, then waits for your review" is a much easier proposition to take to a compliance team than an agent with write access.
Do not use it expecting a hosted service. There is no SLA, no support commitment and no maintenance. The repository says it plainly: "This is a reference implementation; it is not maintained and does not accept contributions."
Do not use it as a consumer shopping tool. Nothing in Claude Commerce gives an individual a shopping assistant in the Claude app; the blueprint is for businesses building agents for their own customers.
And do not deploy the demo. The next section is about why that one matters more than it sounds.
What Are the Limitations of Claude Commerce?
The demo code is not production code, and Anthropic says so without softening it. The solutions FAQ states that "Demo backends are not production hardened," and the repository is more specific about how: "the examples have no authentication and the MCP servers bind to loopback."
Loopback binding and no authentication are not shortcomings to work around later. They are the reason the examples are safe to run on a laptop and unsafe to expose, and anyone standing this up on a public host without replacing that layer has misunderstood what they downloaded.
Everything that makes commerce commerce is yours. "Business rules, authorization, and compliance are the deployment's," says the repository, which is a short sentence carrying a large amount of work.
The sample data is fiction, which is easy to forget once an agent is answering fluently about it. "Every company, brand, product, and person here is fictional; the only company is ACME."
On safety, the architecture guidance makes a point worth carrying into any agent project: "The prompt is where safe behavior starts, but in commerce it can't be where safety is enforced." Enforcement belongs in the interface, not in the instructions, which is the same conclusion the missing charge method expresses in code.
The blueprint does acknowledge real commerce edge cases, including purchase caps. "Most commerce surfaces cap how many of an item one user can buy," the architecture article notes, citing ticket allocations, promotional pricing and fraud control. The solutions FAQ adds that "Cart and refund authority have hard limits" and that anything beyond them escalates to a person.
Now the absences, which on this topic are unusually informative.
Anthropic publishes no pricing for any of this. Not a license fee, not a per-transaction cost, not a modeled example of what running an agent per shopper costs. The only route stated is to "Contact our sales team." Your real cost is ordinary model usage on whichever surface you choose, and you will have to estimate it yourself.
There is no stated geography or region restriction for the blueprint anywhere on these pages, which is not the same as there being none. Availability is governed by the Claude surface you build on, not by the blueprint.
Anthropic does not name any commerce standard. The phrase "Agentic Commerce Protocol" appears nowhere on the pages we read. MCP is mentioned, in the server-side role described earlier. The only other acronym is "UCP," and it appears solely inside a quote attributed to Shopify, where it is never expanded or defined. We are not going to guess what it stands for, and you should be suspicious of any article that does.
The headline conversion statistics carry no methodology. Anthropic's own pages cite figures for larger carts and higher purchase-completion rates, and a partner-attributed conversion figure beside a customer logo, with no sample, period or test design stated for any of them. Treat them as vendor claims, because that is what they are.
And the pages have aged without being updated. The launch article says "The blueprint is available today" and is dated September 2, 2026, with no updated date. More tellingly, the marketing copy is built around getting live before a holiday code freeze, which is simply stale advice in the middle of November. Two smaller inconsistencies point the same way: the article labels Managed Agents as beta while the solutions page omits the label, and the article's own link to its deep-dive points at a path that has since moved.
Before you point an agent at your catalog, make sure your site and data are telling the truth. Get a free audit.
How Does Claude Commerce Compare to the Alternatives?
Against building from scratch, the blueprint wins on architecture and loses on support. You are getting a well-reasoned separation of concerns, two working agent designs and a safety model expressed in code rather than in prose. You are not getting anything you can escalate to. For a team that would otherwise spend a month arguing about where authorization lives, that trade is usually worth it.
Against a hosted agentic-commerce product, the comparison is not close in either direction, because they are different purchases. A vendor platform gives you an SLA, an operator and a roadmap, and takes architectural decisions away from you. Claude Commerce gives you the decisions and none of the operations. Choose the platform if you want someone accountable; choose the blueprint if you want control and already have the engineering capacity to use it.
Against a general-purpose agent framework, the difference is domain judgment. Any agent framework will let you wire a model to a catalog. What the blueprint adds is the commerce-specific reasoning: no charge method in the interface, staged merchant changes, purchase caps as a first-class concern, escalation past hard limits. Those are the conclusions you would otherwise reach by shipping something regrettable first.
And against the thing most people actually imagined when they searched for this, there is no comparison to make, because it does not exist. Claude has no consumer shopping surface. If you want an AI agent that buys things on your behalf as an individual, Claude Commerce is not it, and no Anthropic page claims otherwise.
The honest verdict: Claude Commerce is one of the more useful things a model vendor has published in 2026, and it is useful mainly as an architecture document that happens to compile. Read it for the boundaries it draws. Deploy it only after you have replaced everything it tells you to replace.
Agents are only as good as the data behind them. Request your free audit before you build on top of your current setup.
Final Thoughts
The single most valuable sentence about Claude Commerce is one of the least quoted: "the backend interface the agent calls has no charge method at all." That is what responsible agent design looks like in practice. Not a model instructed to be careful, but an interface in which the dangerous action is simply absent.
Everything else follows from that. The shopping agent discovers and assembles and then hands off. The merchant agent drafts and waits. Payment stays with your checkout or your payments provider. A person presses the button.
So set your expectations accordingly. This is not a feature you turn on, it is not maintained, it has no price list and it ships with authentication deliberately missing. It is a blueprint, in the full sense of the word, and Anthropic has been unusually honest about saying so.
If you want the wider context on how Anthropic uses the word agent across its products, and the four different things people mean by it, our guide to Claude AI agents covers the rest of the picture.
One last caution. Almost everything written about Claude Commerce elsewhere describes it as a shopping product and implies Claude completes purchases. Anthropic's own pages say the opposite, four separate times, in four separate places. When a vendor is clearer than its coverage, read the vendor.
Want this done for you?
SEO Stuff gets businesses cited by ChatGPT, Gemini, Perplexity and Claude, and ranking in Google. Start with the free audit, a call, or the package.
See where you stand
A free, manual SEO + AI search audit of your site, with a report within 48 hours. No credit card, no sales call.
Get my free auditTalk it through
A free 15-minute call with the founder about whether AI search is a meaningful opportunity for your business.
Free strategy callHave it done
The Done-For-You Package: audit, 10 pages of content, 3 DR50+ placements, dashboard. $999 once, delivered in 21 business days.
See the package

