An AI that shops for you? That stopped being a sci-fi demo earlier this week. Claude commerce agents, an open-source blueprint from Anthropic, make agentic shopping real, and the code is free for anyone to use. Here’s what these agents actually do and why Visa and Mastercard are already paying attention.
An AI that shops for you? It’s closer than you think
Picture this: you open a retailer’s app and type “find me a winter jacket under $150 that matches the pair of boots I bought last year.” Instead of scrolling through 40 tabs, the app’s AI searches the catalog, checks your order history, compares three options, and shows you the best pick with a cart ready to go.
That’s the experience Anthropic is trying to make standard. In early September 2026, the company announced Claude commerce agents and released the commerce-agents repository, an open-source reference project that any business can copy and adapt. Claude commerce agents ship with two ready-made agents, four industry examples, and clear rules for keeping money and data safe. And because it’s Apache 2.0 licensed, no one has to pay Anthropic to use it, as the official announcement explains.
What Anthropic just released (in plain English)
The best way to think about Claude commerce agents is as a starter kit for building shopping AI. It’s not a store you visit. It’s the plumbing that lets OTHER companies add AI shopping to their own apps and sites.
The kit includes two distinct agents, and they serve different people.
The shopping agent
This one works inside a merchant’s app and helps customers. It searches the product catalog, handles requests with multiple items, compares options, builds a cart, and answers order and returns questions, all in the same conversation. Its five skills cover search, purchase research, planning goals, customer care, and personalization.
The merchant agent
This one helps the store’s staff instead. It answers sales performance questions, flags low inventory, suggests pricing and promotions, and drafts marketing campaigns. Same underlying system, completely different job.
| What it does | Shopping agent | Merchant agent |
|---|---|---|
| Who it helps | Customers | Store staff |
| Main jobs | Search, compare, build cart, answer orders | Sales insights, inventory alerts, pricing ideas |
| Example request | “Find a flight and hotel under $900” | “Which products are selling slowest this month?” |
| Key skills | Search, research, planning, care, memory | Insights, listings, inventory, pricing, campaigns |
How a Claude commerce agent actually shops
The clever part is how these agents are built. Instead of one giant robot brain trying to do everything, each agent uses something Anthropic calls skills. Think of skills as a toolbox the agent pulls from depending on the task. Researching a purchase? It uses the research skill. Handling a return? The customer-care skill kicks in.
Anthropic says this design beats both of the older approaches: the single mega-prompt that tries to handle everything, and the subagent approach that spawns a different bot for each task. Their tests showed one agent with skills won on quality, cost, and speed. That matters for shoppers because it means responses come back faster without the awkward “handing you over to another department” energy.
The agents even render their answers as interactive components instead of plain text. Ask about hotels and you might get a comparison card you can flip through, not a wall of text. The interface pieces are treated as tools with validation, so the AI can’t invent a price or show you a room that doesn’t exist.
The safety net: approval gates
Here’s the part that should reassure you. In these agents, anything involving money, file changes, or personal IDs is gated in code. The phrase Anthropic uses is “the model proposes, the harness applies.” In practice, that means the AI can suggest a purchase, but the actual checkout step requires explicit approval.
That’s a big deal. It’s the difference between an AI that charges your card and an AI that prepares everything for your one-click go-ahead. Early AI shopping tools that skipped this step are exactly why people worry about agents running wild. Anthropic put the brakes in from day one.
Why Visa and Mastercard are paying attention
The payments giants didn’t stay quiet about this launch. Both Visa and Mastercard were involved in the announcement week, and payments industry outlets flagged the blueprint as a potential standard for how AI agents handle commerce.
The reason is straightforward: if AI agents are going to buy things, the payment industry needs a consistent, safe way for that to happen. A widely adopted open-source blueprint gives them exactly that. When agents from different companies speak the same language around approvals and transactions, the whole system gets safer and faster.
For regular people, this is the story to watch. It means AI shopping won’t be locked inside one company’s app. It’s becoming an open standard that any store can adopt.
What this means for you (shopper and seller)
As a shopper, the near-term change is subtle. You won’t log into “Claude” to shop. You’ll use your favorite stores’ apps, and some of them will quietly get smarter, with agents that remember your preferences and handle the tedious research part of buying.
As a seller, the change is more urgent. The blueprint includes a merchant agent that can monitor your inventory, suggest promotions, and draft campaigns. Small businesses that adopt these tools early could gain the same kind of operational help that used to require a full marketing team.
There’s also a Claude Code plugin called commerce-builder that scaffolds a custom shopping agent with a single command. You don’t need to be a developer to understand what that means: the barrier to building your own commerce AI just dropped dramatically.
The honest caveats
Let’s not oversell it. Claude commerce agents are a blueprint, not a finished shopping experience. Most stores haven’t adopted them yet. The rollout will take months, and early versions will be clunky.
There’s also a trust question. A shopping agent that knows your order history and preferences is convenient, but it’s also a lot of data in one place. The approval gates help, and Anthropic says the agents retrieve only the minimum info needed, but you should still think before granting any AI permanent access to your payment habits.
And remember the broader context: this is the same week OpenAI’s ChatGPT wallet got attention for AI spending money. Agents are coming from every direction, and the industry is still figuring out the rules.
What to do now
Claude commerce agents matter because they turn “AI shopping” from a vague promise into open-source code anyone can run. The shopping and merchant agent pair, the safety gates, and the Visa and Mastercard involvement all point one direction: this is how buying things will work, eventually.
You don’t need to do anything today. But next time an app offers to “find it for you,” you’ll know exactly what’s happening under the hood. And if you run a business, this blueprint is worth a serious look.
Curious about the wider world of AI agents? We’ve got a beginner’s guide to agentic AI vs generative AI and an explainer on AI browser agents. For the money side, here’s what we know about the ChatGPT wallet. And for the safety side of agents, our guardrails guide is a good place to start.
