Anthropic open-sourced a working shopping agent and merchant agent on September 2, 2026, with Shopify shipping its own implementation the same day. The code is free. The catalog quality that determines whether the agent performs is the part most Shopify merchants have not built.
The agent layer just became a weekend project. The catalog underneath it is still the eighteen month job nobody wants to fund.
For most of 2026, the agentic commerce conversation has pointed one direction. ChatGPT, Google AI Mode, and Microsoft Copilot go looking for products, and your job is to be findable when they do. Shopify made that automatic for US merchants on March 24, 2026, with Agentic Storefronts. The work left to you was catalog hygiene, and every readiness checklist published since has been a variation on that theme.
On September 2, Anthropic pointed the arrow the other way. It published a complete reference implementation for two agents you run yourself: a shopping agent that lives on your storefront and a merchant agent that sits in your back office. Apache 2.0 licensed, forkable, with eight working demos across retail, travel, telecom, and ticketing. Shopify published its own implementation the same day, connecting a storefront agent to a live store through the Universal Commerce Protocol.
I have watched enough infrastructure get commoditized to recognize the shape of this one. Something that cost real money last Tuesday is free this Tuesday, everyone reads that as the barrier falling, and the actual barrier turns out to be sitting one layer down where it always was. That is the part worth twenty minutes of your attention, and it is not the part the announcement is about.
Anthropic released two reference agents plus the architecture behind them, not a hosted product you buy. The announcement on the Claude blog covers the harness design, the latency and cost techniques, and the evaluation practices, while the public repository holds the runnable code under an Apache 2.0 licence.
The shopping agent searches a catalog, compares options, handles constraints like under $200 and ships by Friday, builds a multi item cart, answers order and policy questions, and remembers what a customer told it earlier in the conversation. It hands the cart to your own checkout. There is no payment processing in the code at all, which is a deliberate boundary rather than an omission.
The merchant agent is the half that has been undersold. It tracks sales, monitors stock, flags listings running low, proposes markdowns, and drafts a restock plan or a promotion when asked. Every one of those actions generates a proposed change that a human approves before anything moves. That approval requirement is enforced in the code, not in a system prompt, which matters more than it may seem when deciding whether to point something at live inventory during peak season.
Deployment runs through the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI, and there is a Claude Code plugin that scaffolds an implementation against an existing backend. Anthropic names Shopify and Priceline as platforms already running consumer-facing agents on the infrastructure, with Visa and Mastercard working on payment network alignment. The companion architecture guide is the more useful document if you are technical, because it explains the evaluation harness rather than the feature list.
There are two directions in agentic commerce and they require entirely different work from you, which is why merchants keep conflating them and reaching the wrong budget conclusion. The inbound direction is when an agent who belongs to someone else finds your products. The outbound direction is an agent that belongs to you serving a shopper who has already arrived.
You have almost certainly already done the inbound work, or had it done for you. Agentic Storefronts turned on by default for eligible US merchants and Shopify Catalog handles the syndication. The remaining task there is verification and catalog quality, which is what the Agentic Storefronts setup and audit walkthrough covers, and it takes under thirty minutes to confirm. The signals that decide whether an external agent recommends you are covered in more depth in how AI shopping agents actually rank and select products.
The outbound direction is new and it is what Anthropic just released. This is the agent on your product listing page, in place of or alongside your search bar, talking to a shopper who is already on your site. Nothing Shopify has shipped by default gives you this. Until September 2 you either bought it from a vendor or built it, and building it meant a genuine engineering project rather than a sprint.
The distinction decides your budget. If your problem is that ChatGPT does not surface you for your category, a storefront agent solves nothing, because the shopper never reaches your site to use it. If your problem is that a large share of your on site search sessions end without a click, which is a number you can pull from your own analytics in ten minutes, the storefront agent is pointed at a real leak. Diagnose which one you actually have before you fund either. For the wider protocol landscape, including how the Universal Commerce Protocol and OpenAI’s Agent Commerce Protocol relate to each other, the complete 2026 guide to agentic commerce for Shopify merchants is the map.
Anthropic reports that retailers running shopping agents on Claude have seen carts up to 35 percent larger and shoppers 60 percent more likely to complete a purchase, and both figures deserve the same scrutiny you would give any vendor benchmark. They come from Anthropic’s own retail partners. They are not independently audited. And the phrase doing the heaviest lifting in that sentence is “up to,” which means these are the best results observed rather than the median.
I applied the same discipline to the 2026 BFCM data report published by Tapcart last week, where a company that sells mobile apps to Shopify brands concluded that you should build a mobile app. The finding was not wrong. The correct reading was to weight the sections where an independent partner’s transaction data confirmed the survey and to discount the sections where it did not. The same rule applies here, and it applies to a company I would otherwise be inclined to take at its word.
What makes the number unreliable as a forecast is not dishonesty. It is that cart size and completion rate lift from any conversational commerce system is almost entirely a function of catalog quality and integration depth. A brand with clean structured attributes on 4,000 SKUs and a well instrumented cart will see a result. A brand with 4,000 SKUs where the size chart lives in a JPEG and half the variants are named “Option 1” will see the agent confidently recommend the wrong product, which is worse than no agent at all.
The honest way to use these figures is as evidence that the ceiling is real and the architecture works, not as a number you put in a forecast. If you deploy, instrument your own cart size and completion rate from day one so you have a baseline that belongs to you. That is the same measurement discipline that separates merchants who know what their AI channel is worth from merchants who are guessing, and most are still guessing.
A shopping agent is only as good as the product data it reasons over, which means the free code has removed the cheapest part of the project and left the expensive part exactly where it was. This is the load-bearing claim of this piece, and I will defend it plainly: if your catalog is not machine-readable today, a free agent does not move you forward by a single week.
Here is what the agent needs before it can be useful. Product titles that describe the product rather than gesture at it. Variants with real attribute names, so that color, size, and material are separate structured fields instead of a concatenated string. Materials, dimensions, weight, and care instructions as data rather than as sentences buried in a description. Stock status that is accurate to the minute. A returns policy, a shipping policy, and a warranty policy the agent can quote without inventing anything. Compatibility and fit information, if you sell anything where fit is a purchase blocker.
That work does not take a weekend. Across the merchant accounts I worked with during my six years at Shopify, the pattern was consistent enough to be predictable: the catalog is always worse than the founder believes, and the gap is widest in exactly the brands that have grown fastest, because nobody stops to restructure product data while orders are climbing. A team doing this properly on a few thousand SKUs is looking at six to twelve weeks, and most of it is judgment work rather than a script.
Fred Cohen of Quickfire Digital put it about as directly as possible in the BFCM report: LLMs want structure, not just marketing text. That was true when the only agents reading your catalog belonged to OpenAI and Google. It is more true now that you might run one yourself, because an external agent that cannot parse you simply moves on, while your own agent stays on your page and gets it wrong in front of your customer.
The category under real pressure is on site conversational search and product discovery, where the core product is now available as free reference code from the model provider itself. Scot Wingo, who founded ChannelAdvisor and now writes a detailed breakdown of the release and what it means for retailers, counts roughly 200 companies building retailer shopper agents and reads this as genuinely disruptive to that field. I think he is right about the direction and early about the timing.
Apps selling a conversational layer over your catalog now have to justify a subscription against something forkable. Rep AI, Zoovu, Algolia, and Searchspring all sit somewhere in that blast radius, though the ones with deep merchandising rules, tuned relevance models, and years of query data have more defence than a thin chat wrapper does. The vendors most exposed are the ones whose entire pitch was that building this yourself was hard.
The categories this does not touch are worth naming, because the temptation after an announcement like this is to assume everything is now free. Attribution and analytics are untouched, and a tool like Triple Whale, whose Moby agents work across your ad channels rather than your storefront, is solving a different problem entirely. Lifecycle marketing is untouched, and Klaviyo’s position on the retention side does not weaken because a storefront agent exists. Fulfilment, returns, loyalty, and subscriptions are all unaffected.
For a merchant, the practical consequence is not that you should cancel anything this month. It is that your next renewal conversation with a discovery vendor has a new anchor in it. If you are paying $1,200 a month for on site AI search, the reasonable question in that renewal is what the vendor provides beyond what a forked repository and two weeks of engineering would give you. Some will have a good answer. Ask anyway.
The code is free and the running cost is not, which is where most merchants evaluating this will get the math wrong. Model inference is billed per token however you access Claude, and a shopping agent that handles a genuine conversation, retrieves catalog data, and reasons over comparisons consumes considerably more tokens per session than a search query does.
Then there is the integration. Somebody has to implement the storefront backend interface over your catalog, cart, order, and policy systems. The Claude Code plugin scaffolds the project against your stack, which compresses what used to be weeks into something closer to days for a competent developer, but days of a competent developer is still a real number and the scaffold is a starting point rather than a finished deployment.
Then there is the part nobody budgets for, which is evaluation and maintenance. An agent that gives a customer wrong sizing information or quotes a return window you retired last year is a support cost and a trust cost at the same time. The repository ships evaluation harnesses precisely because this is the failure mode, and running those evals is ongoing work rather than a launch task. Add catalog remediation on top and a realistic first year for a mid market brand lands somewhere in the low five figures rather than at zero.
The pattern I have watched break brands between $500K and $2M is premature complexity: a new channel, four new apps, and a build project stacked on top of fundamentals that were never solid. A free storefront agent is an unusually well disguised version of that trap, because the price tag on the visible component genuinely is zero. Below about $1M a year, or with a catalog under a few hundred SKUs where a good filtered collection page does the same job, this is a distraction. Fix the catalog. The catalog work pays off in the inbound direction whether or not you ever deploy an agent of your own.
The highest return move available to you this month is a structured data pass on your top twenty revenue SKUs, not an agent deployment, and that holds at every stage below eight figures. Peak season starts in October for roughly 80 percent of shoppers, and a catalog fix compounds across every AI surface you appear on while an agent build competes for the same engineering hours as your promotional calendar.
The sequence I would run is short enough to start today. Open the self-guided demo before you read another analysis of it, because ten minutes with the working thing tells you more than any write-up, including this one. Export your top twenty SKUs and read them the way an agent would, which means asking whether every attribute a buyer needs exists as a field rather than as prose. Fix what is missing. Then, and only then, decide whether the agent is a Q1 project or a vendor renewal conversation.
Here is the prediction I am willing to be wrong about in public. Twelve months from now, the merchants who deployed a storefront agent on a messy catalog will have quietly turned it off, and the merchants who spent this October restructuring product data will be running the same free code successfully and looking like they had better technology. They will not have had better technology. They will have had a catalog the technology could actually read.
Claude Commerce Agents are an open source reference blueprint published by Anthropic on September 2, 2026 for building two things: a shopping agent that runs on a merchant’s own storefront and a merchant agent that handles back office tasks like sales monitoring, inventory tracking, pricing recommendations, and campaign drafting. The code is licensed under Apache 2.0 and is free to fork from the public GitHub repository. Running it is not free, because the agents consume Claude model usage billed per token through the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI. Budget for inference, integration engineering, and ongoing evaluation rather than for a licence fee.
No. Agentic Storefronts makes your products discoverable to external AI agents such as ChatGPT, Google AI Mode, and Microsoft Copilot, and it has been active by default for eligible US Shopify merchants since March 24, 2026. Claude Commerce Agents point the opposite direction: they are an agent you host on your own storefront, serving a shopper who has already arrived on your site. Shopify published its own implementation of the Anthropic blueprint on September 2, 2026, connecting a storefront agent to a live store through the Universal Commerce Protocol. The two are complementary, they solve different problems, and they require different work from you.
Yes, this requires real engineering capacity and is not a no code install from the Shopify App Store. Someone has to implement the storefront backend interface over your catalog, cart, order, and policy systems, then run and maintain the evaluation harnesses that catch the agent giving customers wrong sizing, stock, or policy information. Anthropic ships a Claude Code plugin that scaffolds the project against an existing backend, which compresses the initial build from weeks to days for a competent developer, but the scaffold is a starting point rather than a finished deployment. Merchants without in house or agency engineering should treat a vendor product as the realistic path.
Anthropic reports carts up to 35 percent larger and shoppers 60 percent more likely to complete a purchase, but those figures come from its own retail partners, are not independently audited, and the phrase “up to” means they are best case rather than median results. Cart size and completion lift from any conversational commerce system depends almost entirely on catalog quality and integration depth. An agent reasoning over vague titles, unstructured variants, and stale stock data will confidently recommend the wrong product, which costs more than no agent at all. Instrument your own cart size and completion rate from day one so you have a real baseline.
No. Below roughly $1M in annual revenue, or with a catalog small enough that a well built filtered collection page answers the same questions, a storefront agent is premature complexity and the same hours spent on product data structure will return more. The catalog work pays off regardless, because clean structured attributes improve how external AI surfaces such as ChatGPT and Google AI Mode recommend you whether or not you ever deploy an agent of your own. Fix titles, variant attributes, materials, dimensions, stock accuracy, and policy pages first, then revisit the agent decision above $2M.
On site conversational search and product discovery apps face the most direct pressure, because their core product is now available as free reference code from the model provider. Vendors in that space including Rep AI, Zoovu, Algolia, and Searchspring will need to justify subscription pricing against a forkable alternative, though those with deep merchandising rules, tuned relevance models, and years of query data have stronger defences than a thin chat wrapper. Attribution platforms such as Triple Whale, lifecycle tools such as Klaviyo, and the fulfilment, returns, loyalty, and subscription categories are unaffected. The practical move is to raise the question at your next discovery vendor renewal, not to cancel anything this month.
