Claude Commerce: How Claude Shops and How Brands Get In

Claude reaches your products through tools, not a crawled index. Anthropic's 2 September 2026 commerce blueprint ships a shopping agent and a merchant agent and leaves payment to you. Here is how that works for a brand, what is documented, what is observed, and what to build first.

Claude commerce is shopping done through Anthropic's Claude. A shopper asks, Claude calls tools exposed over the Model Context Protocol to read a catalog, and a merchant's own agent, built on Anthropic's commerce blueprint, finds products and builds the cart. Payment stays with the merchant (Anthropic, 2 September 2026).

Last updated: 13 October 2026.

What Claude commerce is

Unlike ChatGPT, which ingests a product feed and renders checkout in its own UI, or Google, which runs checkout in AI Mode on a Merchant Center feed, Claude has no merchant feed and no in-chat checkout of its own as of 13 October 2026. It reaches commerce two ways:

• Through tools a shopper's Claude can call. The Model Context Protocol, which Anthropic released on 25 November 2024 as "a new standard for connecting AI assistants to the systems where data lives" (Anthropic, 2024), lets a brand expose its catalog as callable functions. A Claude with that connection can query price, stock and attributes directly. Plain-English version: MCP in the glossary.
• Through a merchant's own agent built on Claude. On 2 September 2026 Anthropic launched "a blueprint to help build commerce agents on Claude", with "the harnesses, patterns, and guardrails an engineering team needs to get a commerce agent running in days" (Anthropic, 2 September 2026). Shopify's framing: merchants "can use Claude to build agents that help customers find products, check out, and answer questions about their orders" (Shopify, quoted by Anthropic, 2 September 2026).

The two meet at the same place: the structured catalog. Where Claude sits among the six platform assistants is in the six AI agents compared; the on-site family it can also power is in AI shopping assistants: the 2026 map.

How Claude shops: tools, not pages

This is the part that changes the work. A search assistant reads your product page. A tool-calling agent reads the tool's response.

After reading Anthropic's reference repository, ReFiBuy's Vedant Gupta followed one shopper sentence through the consumer agent: "What the agent gets is a short row of fields from the catalog", and the product page itself "does not really come into it" (Vedant Gupta via Scot Wingo, 11 September 2026). The tool response is the product, as far as the agent is concerned.

Three consequences:

1. Fields decide, prose does not. If size, material, pack count, compatibility and live availability are not fields the tool returns, the agent cannot filter on them, and a constraint it cannot filter on drops you before price is compared.
2. Comparison happens inside the agent. A shopper asking Claude to compare two options gets the comparison built from the fields it was given. Honest, specific attributes serve you; marketing adjectives are not fields.
3. One endpoint serves many clients. An MCP server is not Claude-specific. The same connection is readable by other MCP-speaking agents, which is why the protocol is the connective layer under ACP and UCP rather than a rival to them; see ACP, UCP, AP2 and MCP explained.

What the blueprint ships, and what it leaves to you

Anthropic's release contains reference implementations of a shopping agent and a merchant agent across retail, travel, telecom and ticketing, and a Claude Code plugin (Anthropic, 2 September 2026). Retailgentic's read of the repository: two open-source reference agents with eight reference examples across retail and marketplaces, travel, telecom and entertainment, Shopify as a launch partner, and a blueprint that ships the conversation and the cart while checkout, fraud rules and protocol coverage are still the merchant's to add (Retailgentic, 2 September 2026).

Anthropic says the same about payment: the blueprint "leaves payment to you, whether that is your existing checkout or an agentic payments provider" (Anthropic, 2 September 2026). Visa and Mastercard are named as collaborators on that payment side. A brand that builds on it owns the order, the fraud rules and the merchant-of-record position, as on every other agent surface; the comparison is in AI shopping agents: how they buy.

Anthropic also makes a performance claim: retailers running shopping agents on Claude "have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase" (Anthropic, 2 September 2026). It is Anthropic's number about its own customers; no retailer, baseline or window is named, so treat it as a vendor claim, not a benchmark.

Who is already moving

• Retailers. In ReFiBuy's Q2 AI1000, co-developed with Digital Commerce 360, the number of Top 1000 retailers whose largest AI traffic source is Claude rose from one to 15 in a quarter; over the same period ChatGPT-led retailers fell from 844 to 722 (ReFiBuy, 18 August 2026).
• Referral share. Retailgentic's read of Similarweb's referral data: "Claude hasn't surpassed Perplexity in our dataset, but one more Q like this and that will happen" (Retailgentic, 18 August 2026).
• Merchant tooling. On 17 September 2026 ReFiBuy announced a bi-directional connection for the Claude commerce agent: enriched product data in, anonymised shopper context back (ReFiBuy, 17 September 2026).

What to do this quarter

1. Audit the fields, not the copy. For your top SKUs, list which attributes exist as structured data with live availability at variant level, and which exist only in prose or images. The second list is what Claude cannot see. The general method is in optimising your catalog for AI agents.
2. Expose the catalog over MCP. Stand up an MCP server that returns product, variant, price, stock and the attributes shoppers filter on. This is the single integration that serves Claude and every other MCP-speaking agent; our own is documented at connect Tru Commerce over MCP, and securing one is covered in MCP security.
3. Decide whether to build your own agent. By Anthropic's own description the blueprint gets "a commerce agent running in days" (Anthropic, 2 September 2026). It fits complex catalogs where a conversational interface earns its place; for simple ranges, the MCP endpoint alone does the work.
4. Make the checkout finishable. Whichever agent arrives, it needs a guest path, no captcha at the pay step and fraud rules that do not auto-decline a fresh card. Enable agentic checkout is the product we run this on.
5. Fix attribution before you judge the channel. Agent orders land in GA4 as Direct. Scot Wingo of ReFiBuy coined the term Dark Agentic Commerce Traffic for it; for MyMuse, GA4 showed ₹11.5K/mo of AI revenue against ₹81.2K/mo actual, 7× hidden (MyMuse case study). Method: dark agentic commerce traffic.

Ads and ranking: what Anthropic has said

On 4 February 2026 Anthropic wrote: "Claude will remain ad-free." Users "won't see 'sponsored' links adjacent to their conversations with Claude; nor will Claude's responses be influenced by advertisers or include third-party product placements our users did not ask for" (Anthropic, 4 February 2026). There is no sponsored slot to buy and, for now, no paid shortcut into a Claude answer.

Nor is there a published ranking rule. OpenAI publishes one for ChatGPT ("availability, price, quality, and whether they are the maker or primary seller", OpenAI Help Center, 2026); Anthropic publishes none for Claude. What a merchant-built agent recommends is whatever its tools return and its instructions say.

Where Claude fits in your surface strategy

• Technical, prosumer and research-heavy categories: Claude early, alongside Perplexity. The shopper is comparing, and comparison runs on fields.
• Complex catalogs: a brand-owned agent on the blueprint is worth prototyping.
• Mainstream consumer and commodity ranges: cover it with the MCP endpoint you build anyway, and concentrate effort on ChatGPT and Gemini, where the published entry paths are feeds.

Surface page: Claude for brands.

Related: the three September posts

On 2 September 2026 Anthropic published an Apache 2.0 blueprint for a shopping agent a business embeds in its own app and a merchant agent its staff use; it ships the conversation and the cart, and leaves payments, ranking and authorization to the deployment (Anthropic, 2 September 2026). The cluster: Claude for Commerce explained, what a merchant has to build and how a Claude shopping agent picks products.

Where Tru Commerce fits

MCP App Builder deploys a brand's catalog as a live MCP server that agents can query and buy from. On Provogue the endpoint is live, so the product data the on-site assistant reads is readable by off-site agents on the same terms: one enrichment, two surfaces. We have run this on Gray Porter and a group of US DTC brands: we made the product pages and checkout readable and finishable by an agent, then placed a real Muse order on the store. To see what Claude would get back from your catalog today, book a demo.