Claude commerce agents: what a merchant has to build

Anthropic's blueprint of 2 September 2026 gives a merchant's engineers two reference agents, a backend contract and a checkout tool whose URL the model never sees. It gives them no payments, no ranking, no user auth and no directory. Here is the split, the three runtime paths, where MCP enters, and what to build first.

Claude commerce agents are the two reference agents in Anthropic's open-source blueprint (2 September 2026): a shopping agent you embed in your own app and a merchant agent your staff use. You implement backend, checkout and payments; Anthropic ships neither ranking nor storefront.

Last updated: 12 October 2026.

What the blueprint gives you

Anthropic calls it "the harnesses, patterns, and guardrails an engineering team needs to get a commerce agent running in days" (Anthropic, 2 September 2026). What shipped and what did not is in Claude for Commerce explained; this post reads the repository for the engineer who has to build on it.

The shopping agent. "The shopping agent lives inside your app or website" (Anthropic, 2 September 2026). It searches the catalog, handles multi-item requests, remembers preferences, shows products and the cart in the conversation, hands the cart to checkout and answers order and policy questions, under "guardrails designed to constrain prices and products to actual catalog data".

The merchant agent. It reads what is selling, flags inventory, recommends pricing from the store's own history and drafts campaigns. Nothing goes live on its own: "every write is a staged change the host's approval surface applies", and "A preview card approves nothing; an approval typed in chat sets nothing" (Anthropic, README).

The StorefrontBackend interface. The contract your deployment implements: "11 required methods, plus get_account_context, get_disclosure, and checkout_handoff" (Anthropic, core README). Search is one of them, and the model reads your results as compact rows: "A compact row is 70 to 120 characters, so a family holds up to about sixty variants" under a 12,000-character cap (Anthropic, docs/backends.md). Vedant Gupta's reading: the agent gets "a short row of fields from the catalog" and "the product page itself does not really come into it" (via Scot Wingo, 11 September 2026).

Server-issued product ids. "Cart writes accept only product ids a catalog or order tool returned this session, or lines already in the cart" (Anthropic, README), with a per-item cap, a line-count cap and serialized writes. The model cannot invent a SKU.

The checkout tool. checkout "renders the cart for the host to complete". "The checkout card links to your own checkout route, or to the platform's hosted checkout URL (one per seller on a marketplace). The backend returns the URL and the host renders it; the model never sees it" (Anthropic, README). StorefrontBackend has no payment method.

The plugin. The commerce-builder Claude Code plugin brings "Six skills and four commands" and "runs no code of its own" (Anthropic, plugin README). Default models are claude-sonnet-5 for shopping and claude-opus-5 for the merchant agent (Anthropic, docs/deployment.md).

What it leaves to you

The README's sentence: "Business rules, authorization, and compliance are the deployment's." docs/safety.md lists what a deployment owns:

Credentials, rate limits and log hygiene are on the same list (Anthropic, docs/safety.md, read 10 October 2026). Web search is "registered only when enable_web_search is set", so by default the agent knows only what your backend returns. What that does to recommendations is in how Claude picks products.

Three runtime paths, and where MCP sits

The repository runs the same prompt, skills and tool contracts three ways (Anthropic, README):

1. Messages API: "The reference loop". Your code runs the loop and calls Python backend methods.
2. Claude Agent SDK: "The same prompt, skills, and tools, with the SDK running the loop". Still backend methods.
3. Claude Managed Agents: "A hosted agent over the same skills and contracts, calling your MCP server".

MCP is one path, not the mechanism. "MCP connectors: None ship; both agents reach your systems through the backend interfaces." A platform's own MCP server for catalog, cart or checkout "is called from a backend method server-side; on Managed Agents the manifest mounts it beside the role's server" (Anthropic, README). The accurate sentence: on the hosted path the agent calls your MCP server; on the API paths it calls backend methods.

The first two paths deploy to "the Claude API, Amazon Bedrock, Microsoft Foundry, or Google Cloud Vertex AI"; Managed Agents runs only on Anthropic infrastructure or Claude Platform on AWS (Anthropic, docs/deployment.md). The Shopify connector in the Claude connectors directory, added April 2026 with 25 admin tools, is unrelated to the blueprint (Anthropic, connectors directory, read 10 October 2026). A catalog MCP server of the kind the third path calls is what MCP App Builder deploys; the authentication the examples leave open is in our MCP security checklist.

Shopify's implementation, the worked example

Shopify/claude-for-commerce-examples was created on 1 September and announced on 2 September 2026 (GitHub API). Its description: "a storefront shopping agent over UCP and Sign in with Shop, and a merchant agent over the Admin API" (Shopify, README).

"Point SHOP_DOMAIN at a store and the agent searches its live catalog, builds a real cart, answers from its policies and FAQs, and hands the customer to the store's own checkout page." SHOP_DOMAIN "is any domain serving /.well-known/ucp". Orders follow "Order MCP v1 semantics"; the merchant side "ships no web UI". And: "complete_checkout is never called - it places the order and takes payment, which this repository never does" (Shopify, README, read 10 October 2026). Vanessa Lee, VP Product at Shopify, calls it "a reference storefront implementation that connects them to a merchant's store through Catalog, UCP and Shop Sign-in" (claude.com, 2 September 2026). The no-code alternative is Shopify Inbox.

Since then: "the UCP path never touches your DOM" (Shopify Community, 3 September 2026); the hosted UCP profile omitted the loyalty extension as of 25 September (one practitioner report); and no completed-order report appeared in the thread by 25 September. The repo had 144 stars and 21 forks on 10 October 2026 (GitHub API).

Against the other developer surfaces

Sources: Anthropic repository, read 10 October 2026; the four shopper-side columns are from AI shopping agents: how they buy. The first column is the only one where you host the agent; the others visit you. Anthropic's code speaks none of ACP, UCP or AP2, so ChatGPT or Gemini checkout is built separately. Meta's connector route is in Muse connectors.

What to implement first

1. Catalog endpoint. Search and product details as compact rows, "70 to 120 characters", with options matched as values so "a filter like size = king works" (Anthropic, docs/backends.md). "Return only fields the model needs"; no "image URLs on every search row" (deep dive, 2 September 2026). A field missing from the row is missing from the answer.
2. Variant-level stock and price. "Storefront price is the lowest in-stock variant's (a 'from' price); a family is in stock while any variant is" (Anthropic, docs/backends.md). Return both per variant, and treat a price "computed per request" as "a quote for that context".
3. Cart with server ids. Accept only ids your catalog or order tools returned this session; enforce the per-item and line-count caps; serialize writes.
4. Checkout hand-off. Implement checkout_handoff to return your route or hosted URL. Payment, 3DS, fraud and eligibility stay on your rails; a guest path and fraud rules that do not auto-decline an agent-started session are the work behind enable agentic checkout.
5. Order record with the agent marker. The host places the order after the hand-off, so only the host knows the session was agent-started. Write that onto the order; the blueprint does not, and without it the order lands in GA4 as Direct. For MyMuse, GA4 showed ₹11.5K/mo of AI revenue against ₹81.2K/mo real, 7× hidden.

Then evals: the deep dive recommends 50 to 100 cases per user flow and reports "90-99% cache hit rates" in the best deployments (Anthropic, 2 September 2026).

The open questions

No directory. No merchant directory, listing programme or Claude storefront result exists to get into; the agent sits in your app and Claude.ai users never meet it. Hacker News noticed: "Anthropic expects that the buying agent will sit inside the merchant's own platform or application" (nandy_tushar, September 2026). The connectors directory takes submissions with no fee or review terms stated, but it lists assistant connectors, not storefronts.

No pricing. Cost is API usage; the post routes to "Contact our sales team". No fee, revenue share, listing or placement programme appears in the primary pages (Anthropic, read 10 October 2026).

No ranking. Anthropic ships none; every "how Claude ranks products" claim is a vendor's reading, named as such in how Claude picks products.

No completed checkout on record. Neither the Shopify Community thread (3 to 25 September 2026) nor Hacker News holds a merchant reporting a finished order through either repository. PYMNTS: Anthropic "built the shopping brain but skipped the wallet entirely" (PYMNTS, 3 September 2026).

The brand-side view is in Claude commerce: how Claude shops and how brands get in; the surface page is Claude for brands.

Where Tru Commerce fits

Our MCP server connects from Claude: as a custom connector in the web and desktop apps, and in Claude Code with claude mcp add; the steps are at connect Tru Commerce from Claude. That hosted server is read-only; the catalog server a brand deploys for its own shopping agent is MCP App Builder. 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 have your catalog and checkout checked the same way before your team builds on the blueprint, book a demo.