Anthropic has published the code for building AI shopping assistants and merchant agents as free open source, a notable move in what the industry calls agentic commerce. On September 2, 2026, the company open-sourced Claude Commerce Agents, a reference blueprint containing two working agents, a customer-facing shopping agent and a staff-facing merchant agent, with complete example implementations for retail, travel, telecom, and ticketing, plus a Claude Code plugin that scaffolds a custom agent against a store’s own backend.
Anthropic says retailers using Claude shopping agents have seen carts up to 35 percent larger and shoppers 60 percent more likely to complete a purchase. Its head of product for the Claude platform, Angela Jiang, told Reuters the cart figure came from one partner, and the company has not named the retailers, sample sizes, or methodology, so treat the numbers as company-reported rather than benchmarks. The code is Apache 2.0 licensed, and a launch testimonial describes getting both agents running locally in under an hour, though running them in production still costs Claude usage, engineering, security, and maintenance. The release lowers the cost of prototyping retailer-controlled agents while leaving companies responsible for their data, integrations, security, and operations. Here is what is in the blueprint, what it asks of a retailer, whether it could pressure specialist vendors, and what it means if you shop or sell online. (Individuals named in this article are referenced for news context only; they are not affiliated with Coursiv and do not endorse it.)
What Anthropic Released
Claude Commerce Agents is not a product you subscribe to. It is a public repository, anthropics/commerce-agents on GitHub, described by Anthropic as the harnesses, patterns, and guardrails an engineering team needs to get a commerce agent running in days rather than months. It ships under the Apache 2.0 license, runs on the Claude API with deployment guides for Amazon Bedrock, Microsoft Foundry, and Google Cloud Vertex AI that vary by runtime, and is explicitly labeled a reference implementation: Anthropic says it will not maintain it as a product or accept outside contributions.
The shopping agent
The customer-facing agent lives inside a store’s app or website. It searches the catalog, assembles multi-item requests such as everything needed for a camping trip, personalizes recommendations from a shopper’s history, shows products and side-by-side comparisons inside the conversation, builds the cart, hands off to the store’s existing checkout, and answers service questions about order status, returns, and refunds. It remembers preferences across sessions, which several commenters singled out.
The merchant agent
The second agent faces staff rather than customers. It explains sales performance and analytics, maintains listings, raises inventory alerts, recommends prices and promotions, and drafts marketing campaigns. Every change it proposes is staged: nothing goes live until a human approves it. The same pattern, an analyst that drafts and a person who decides, also appears in other agent products launched this month.
Four Industries, Four Working Demos
| Vertical | What the demo handles |
|---|---|
| Retail | Search, comparison, cart, checkout handoff, memory across visits |
| Travel | Date-bound inventory, itinerary building, occupancy calendars |
| Telecom | Account context, plan comparison matrices, regulated fees |
| Entertainment and ticketing | Timed seat holds, waitlists, venue maps, fee disclosures |
Each vertical comes with both a storefront and a back-office portal that run locally on fictional data, so teams can watch both agents before wiring in real systems. The travel and ticketing examples matter because they cover the awkward cases, expiring holds and date-specific stock, that a generic shopping demo skips.
The Claude Code Plugin
The fastest path in is a Claude Code plugin called commerce-builder. It adds four commands: one scaffolds a new agent from a plain-language description of the store, one extends an existing agent with a new flow, one writes an evaluation harness, and one reviews an agent already built. A customer testimonial in Anthropic’s launch materials describes running both reference agents locally in well under an hour with live conversations working on the first attempt. Shopify has separately published example implementations that connect the blueprint to a live Shopify catalog and cart through its commerce endpoints; the storefront example still sends shoppers to Shopify’s own checkout, and the merchant example records changes to a ledger that an operator must apply.
How to try the demos
For a team that wants to see it before committing, the path is short: clone the repository, add a Claude API key, install the example web apps, and run one of the four verticals with a single script that launches the storefront, the merchant portal, or both. The demos run on fictional data, process no payments, and bind their servers to the local machine by default.
How the Guardrails Work
Shopping is where an agent can do real damage, and the safety design is where the blueprint is most specific. Anthropic’s engineering guide states that the safety rules it describes are enforced in code on both agents rather than left to model instructions, and that every backend read passes through a sanitizer that strips content imitating tool calls or conversation turns, a defense against prompt injection hidden in product data.
The agent never touches money
Checkout is handed back to the retailer’s existing system or to an agentic payment provider; the checkout URL is never exposed to the model at all. In the demos, no order is ever charged.
Prices and products come only from the catalog
The cart accepts only product IDs the server returned in that session, and product cards render only records the server filled in, so a fabricated listing or price cannot reach the cart or a product card even if the model’s prose is imperfect; Anthropic also says the prompts are designed to avoid manipulative upsell patterns.
Humans approve every merchant change
All merchant-side writes are staged behind an approval gate, provenance checks fence what tools may write, analysis delegates run under budget caps, and per-turn limits on tool use are set in configuration. The repository documents more than fifteen such rules with the exact modules that enforce them.
What the Blueprint Asks of a Retailer
Here is the catch: the blueprint provides interfaces and agent tools for catalog, cart, and checkout, not a retailer’s actual systems or data. Retailers must implement two backend interfaces themselves, one for the storefront and one for merchant operations, and the agent only ever sees structured results from those calls.
A code review by commerce data specialist Commerce Clarity spells out what that demands. In the reference implementation, deterministic filtering works on structured fields, so a condition a shopper states, such as waterproof or under two kilograms, needs to exist as a normalized attribute rather than only in description text. The reviewer found that mixed units across a catalog, grams here and pounds there, caused filters to miss in the demo backend, and that the demo search defaults to eight results and drops a filter that would return nothing rather than telling the shopper. Those are choices in the reference code that a retailer’s own backend can change, but they show where the work lies. The reviewer’s summary: catalog volume used to be the advantage, and an agent rewards correctness instead. For many stores, the real project is not the agent but the data underneath it.
Who Is Behind It, and Who Should Worry
Anthropic’s launch materials name Shopify, Priceline, Accenture, Mastercard, Visa, Intuit, Klaviyo, Wix, Zomato, Fetch, and Square in connection with the launch; the roles differ, from platform partners such as Shopify and payment networks such as Visa and Mastercard to consultancies and customers quoted in testimonials, and Anthropic does not say which of them run the blueprint in production.
Retail commentator Scot Wingo, who also co-founded an agentic commerce startup, framed the release as a shock to the middle of the market: Anthropic is handing out the code for a very powerful shopper agent for free beyond token costs, which democratizes something previously sold by specialist vendors. By his own estimate, offered without a published methodology, roughly 200 companies build retailer shopping agents, and he flagged the Shopify app ecosystem and marketplace management platforms as the most exposed. The winners in his reading are small and mid-sized merchants who could never afford that tooling before. The counterargument is that free reference code does not commoditize a production service: live catalog and inventory integration, identity and permissions, monitoring, compliance, refund workflows, and ongoing maintenance are where specialized vendors earn their keep, and the blueprint supplies none of that.
What the Skeptics Raised
Some users on X questioned the parts the demos do not show. The sharpest technical point: the hard part is staleness, since stock and price can change between the agent reading a page and reaching checkout. Others asked for the flow nobody demos, the return, where money travels backward through the merchant stack and the agent must escalate to a human. A third theme was memory: these agents only become useful when they remember sizes, budgets, and past returns without being retrained every session.
The Bigger Picture: Agentic Commerce Goes Open
Several major technology companies now want a position in how people buy, and most have chosen open protocols too. OpenAI’s in-chat checkout runs on an open-source agentic commerce protocol developed with Stripe that leaves payment, fulfillment, and the customer relationship with the merchant. Google’s Universal Commerce Protocol is likewise open source and built to connect agents to existing retailer and payment systems. Amazon’s Shop Direct surfaces more than 100 million products from over 400,000 outside merchants, and its Buy for Me feature completes purchases on those merchants’ sites. Against that backdrop, Anthropic’s distinction is narrower than the launch chatter suggested: it is not operating a consumer shopping destination at all, but supplying retailer-hosted reference applications and the agent architecture behind them. That lowers the cost of prototyping a store’s own agent; it does not by itself remove the integration, data, safety, and operations work that production requires.
What It Means for You
If you shop online
More stores will likely offer an assistant that turns a request like a gift for a niece who likes building sets under $45 into a real cart and answers where your order is. The useful test is memory and honesty: does it remember you, and does it say when the store lacks what you want.
If you run a store
For a small business, the merchant agent may be the more immediately practical half: a data analyst that explains performance, flags inventory problems, and drafts campaigns, with you approving every change. The prerequisite is the unglamorous one the code review identified, clean and structured product data. Stores with clean, structured catalogs have a shorter path; stores with messy ones have a data project first.
The Skill Underneath Every Agent
Whether you are approving a merchant agent’s price change or asking a shopping agent for exactly what you need, the leverage is in the brief and the review: describing a goal precisely, judging whether the output is right, and knowing where the system’s blind spots are. Coursiv builds those foundations with step-by-step guides, short daily lessons, and hands-on practice with AI tools, designed for busy people without a technical background. Check the official site for current course details and pricing.
What to Watch Next
Watch which platforms productize the blueprint; Shopify has already published examples, and a one-click version would reach millions of stores. Watch for independent numbers on cart size and conversion beyond Anthropic’s own customers. And watch the return and refund flows, which test integration, authority, and escalation to a human, the hard part several users pointed to.
The Bottom Line
Anthropic did not launch a shopping product; it published the plans for one and made them free. That lowers the cost of prototyping a retailer-controlled agent without removing the expensive parts, integration, data quality, safety, and operations, and it pressures vendors selling generic agent scaffolding more than it proves specialized commerce platforms have lost their edge. The open questions are the ones shopping has always had: stale inventory, returns, and whether the assistant remembers you.