Claude Commerce Explained: What Anthropic Shipped, What It Is Not, and How to Run It on Your Platform

“Claude commerce” became one of the fastest-growing search terms in ecommerce this summer. On September 2, 2026, Anthropic gave the term a product. It released Claude Commerce Agents, an open-source blueprint for shopping agents and merchant agents that run on Claude.

Most coverage repeats the launch post. This guide does a different job. It explains what the code does and what it does not do. It maps the blueprint to the platforms we build on at Elogic Commerce. It lists the changes a B2B store must make before it can use it. Every fact links to the primary source.

Key facts about Claude Commerce

• Claude Commerce is the common name for Claude Commerce Agents, an open-source blueprint from Anthropic for shopping agents and merchant agents built on Claude.
• Released on September 2, 2026, under the Apache 2.0 license, in the GitHub repository anthropics/commerce-agents.
• Contains two working agents, four vertical examples (retail, travel, telecom, ticketing), a safety harness, eval patterns, and a Claude Code plugin.
• Does not complete purchases, process payments, or replace your commerce platform. Checkout stays on your store. Every merchant change waits for human approval.
• Reported results: carts up to 35% larger and shoppers 60% more likely to complete a purchase. These are Anthropic’s own figures, not an independent benchmark.
• Runs on the Claude API, Amazon Bedrock, Microsoft Foundry, and Google Cloud Vertex AI. Shopify and Priceline run agents on Claude. Accenture, Mastercard, and Visa are named ecosystem partners.
• Two platform reference builds exist: Shopify (storefront agent over UCP and Sign in with Shop, merchant agent over the Admin API) and Shopware (open source under MIT, runs in Docker or in a browser tab, with a 107-case eval suite).

What is Claude Commerce?

Claude Commerce is the search term. The product name is Claude Commerce Agents. Anthropic also calls it “a blueprint for building commerce agents on Claude”. Some publications call it “Claude for Commerce”. All four names point to the same release of September 2, 2026.

A blueprint is not a hosted product. It is working code, patterns, and guardrails that your engineering team forks and adapts. Anthropic states that the repository is a reference implementation. Anthropic does not maintain it as a product and does not accept outside contributions. Your team owns the result.

What is in the release

The release has three parts.

1. The repository. anthropics/commerce-agents on GitHub, under the Apache 2.0 license. It contains a shopping agent, a merchant agent, four runnable vertical examples, a safety harness, eval patterns, and deployment paths.
2. The engineering guide. “A guide to the anatomy of effective commerce agents” on claude.com. It describes the architecture, the latency and cost techniques, and the eval practices. It is the most useful document in the release.
3. The demos. Live demos for retail, travel, telecom, and ticketing at claude.com/solutions/commerce. You can see both agents work before you write code.

The repository also includes a Claude Code plugin named commerce-builder. The plugin has four commands. /scaffold-commerce-agent builds a new agent from a plain-language description of your store. /add-commerce-flow adds a flow to an existing agent. /author-commerce-evals writes an eval suite. /review-commerce-agent reviews an agent you already have.

The agents are Python 3.11 or later. The example web apps are Node 22. One agent definition (prompt, skills, tool contracts, and gates) runs on three runtimes: the Messages API, the Claude Agent SDK, and Claude Managed Agents (beta).

Who already runs it

The launch post includes statements from Visa, Mastercard, Accenture, Priceline, Intuit, Shopify, Klaviyo, Wix, Zomato, Fetch, and Square. Priceline rebuilt the latest version of its assistant, Penny, on Claude. Wix reports a working commerce agent that took prompts within fifteen minutes. Fetch reports both agents running locally in under an hour.

Two platform vendors published reference implementations within days. Shopify released Shopify/claude-for-commerce-examples. It connects the shopping agent to a Shopify store through Catalog, UCP, and Sign in with Shop, and the merchant agent through the Admin API. This is the reason the query “claude commerce shopify” exists. Shopware’s Agentic Commerce Lab released agentic-commerce-lab/shopware-claude-commerce, which runs the blueprint unmodified against a real Shopware 6.7 shop. Both builds are compared below.

Is Claude Commerce real? Search demand before and after the release

Yes, it is real, and the market expected it. Ahrefs data shows that global monthly searches for “claude commerce” rose from 5 in February 2026 to 4,852 in August 2026. The US volume in August was 900. The rise in June and July came before the release. Anthropic did not announce the product then, so the early demand came from expectation, not from a launch.

Source: Ahrefs Keywords Explorer, exported September 6, 2026.

The related queries show what people want to know: “is claude commerce real”, “claude commerce release”, “claude commerce vs shopify”, “claude commerce platform”, and “how to use claude commerce”. This guide answers each one.

What Claude Commerce is not

Most confusion about Claude Commerce comes from six wrong assumptions. The table corrects each one with the fact from the Anthropic documentation.

The two agents in the blueprint

The blueprint ships two agents. They sit on opposite sides of the transaction. Each one is defined once (prompt, skills, tool contracts, gates) and runs on all three runtimes.

The query “claude e commerce skills” refers to these ten skill names. A skill is a directory with a SKILL.md file. The agent loads it when the conversation needs it. You can add your own skills under the same directory.

How the architecture works: five rules

The engineering guide is written for engineers. These five rules summarize it for a commerce team that must decide what to build.

Rule 1. One agent, not many subagents

One Claude instance runs the full conversation. There is no intent router in front and no set of domain agents behind. Skills give the agent modular knowledge without a handoff. Anthropic reports that across several enterprise deployments, a single agent with skills beat both the one-prompt design and the subagent design on quality, often at lower cost and latency.

Subagents keep two roles. A subagent can run a narrow task with its own context window, such as deep research over documents. A domain that already has its own agent, such as a pharmacy or financial-services flow with its own compliance surface, gets a hand-off. In a hand-off, the domain agent takes over the conversation. In delegation, the main agent stays in charge.

Rule 2. Decide prompt or skill by frequency

Instructions that a third or more of the traffic needs go in the system prompt. The rest goes in skills. Product search is in the prompt, because nearly every session uses it. Safety rules, legal rules, brand rules, and critical user facts such as allergies always go in the prompt. If a skill is predictable from the entry page, the harness injects it before the first model call.

Rule 3. Tools call the systems you already run

Your store already has search and ranking, a cart, a profile store, an inventory system, promotion engines, and sales analytics. The agent’s tools call those systems. They do not rebuild them. When the agent calls search_products, the results arrive ranked by your search engine. The model decides which results serve the goal, how many to show, and how to present them. Tool results return only the fields the model reasons with. Image URLs on every search row are the usual mistake.

Rule 4. UI components are tools

Most commerce answers are components, not prose: a product carousel, an itinerary, a seat map, a chart. The agent calls a presentation tool such as present_products with typed arguments. Your server validates and enriches the call. Your client renders it. Because the call sits in the messages array, the agent knows what is on screen. When a customer says “the third one”, the layout is in the last presentation call.

Rule 5. Safety is enforced in the harness, not in the prompt

A prompt rule is one injection away from being skipped. In commerce, the failures cost money and are often irreversible. The blueprint enforces four rules in code, on both agents, on every runtime.

1. The model stages, a person or a policy applies. No model tool call moves money or changes the business. Order placement, payments, refunds, price changes, and campaign launches end in an action the harness controls.
2. Writes and renders accept only server-issued IDs. The harness records every ID the server gave the model in this session. A product ID that arrived any other way, hallucinated, pasted by a user, or planted in a review, is refused before the backend sees it.
3. Caps hold against repeated requests. Limits are enforced on the resulting state, not on the request. Cart writes are serialized per session, so parallel tool calls cannot stack past a cap. Merchant changes are checked against caps on price movement, discount depth, restock size, and campaign budget.
4. Third-party content is sanitized. Listings, reviews, policies, seller messages, and stored memory pass through one sanitizer and arrive fenced with a fixed label. The prompt treats fenced text as material to report on, never as instructions.

Claude Commerce vs Shopify, OpenAI, and Google

“Claude commerce vs shopify” is a common search. It compares two different things. Shopify is a commerce platform. It holds the catalog, the cart, the checkout, and the payments. Claude Commerce is agent code. It reads from a platform and stages changes for it. A Shopify Plus store can run Claude Commerce, and Shopify built the reference implementation for that.

The useful comparison is between the ways to put an agent in front of shoppers. The table places Claude Commerce next to the three other approaches that matter in September 2026.

The pattern of 2026 is clear. Shoppers discover in AI assistants and buy on the merchant’s property. Claude Commerce is built for the second half of that pattern. It gives the merchant an agent it controls, on its own site, with its own checkout. UCP and ACP cover the first half: they let external agents find your catalog. A mature store will need both.

Two reference builds you can read: Shopify and Shopware

Two platform vendors have implemented the two backend contracts against a real store and published the code. Both leave the blueprint packages untouched and pin them to a commit. Together they show how much work sits between the blueprint and a live shop, and where that work lands on a platform team.

What the Shopware build proves

• The blueprint runs unmodified. The repository installs the blueprint packages at a pinned commit and adds two thin FastAPI hosts, one per contract. A platform team that does the same never forks the agent. It forks the execution layer.
• The platform can own the approval loop. Shopware’s Admin MCP server defaults every write tool to dryRun=true. That gives the stage, preview, approve, apply loop the merchant agent requires, with the preview generated by the shop. The ACL templates split the stager role from the approver role, so the agent’s own role cannot apply its proposals.
• Regulated disclosures are server-authored. Base price (the German PAngV Grundpreis), delivery time, and VAT notice are lines derived from product data, not model output. The blueprint’s telecom example uses the same pattern for fees. EU stores need this on day one.
• Evals ship with the code. 107 cases, 64 shopping and 43 merchant, every positive case paired with a negative one. The CI gate requires a safety pass rate of 1.00, a core pass rate of at least 0.90, a cache hit rate of at least 0.80 from the second trial, and a cost per turn of at most 0.10 USD for shopping and 0.30 USD for merchant turns. The merchant CI set moved from 23 of 38 to 34 of 38 after host prompt rules. That is what a first month looks like.
• The limits are stated. Traffic and conversion are reported as unavailable, because Shopware core does not measure traffic. Campaigns are not applicable in this version. Identity Linking waits for https. The browser demo downloads about 150 MB and boots in 20 to 40 seconds. A reference build that states its limits is more useful than a demo that hides them.

Shopware’s Agentic Commerce Lab describes the split as an intelligence layer and an execution layer. The model reasons. The platform stays the system of record and executes: prices, promotion rules, stock, delivery promises, and the order. Their claim is that the execution layer is the same position whether the customer arrives through Claude, ChatGPT, Gemini, or the merchant’s own agent. That matches the pattern of 2026 above. It also explains why the platform map below matters more than the model choice.

Platform integration map: what your platform must supply

The blueprint ships no platform connectors. Both agents reach your systems through two backend interfaces that your team implements: StorefrontBackend for the shopping agent and MerchantBackend for the merchant agent. Each backend method calls your service on the server side with the credential your host holds for the session. The model reads only the result.

The two tables map each interface to the platforms Elogic Commerce builds on. The mapping is at the API family level. Exact endpoints depend on your platform version and your extensions.

Table A. Shopping agent: StorefrontBackend

Table B. Merchant agent: MerchantBackend

Two notes apply to every row. First, the repository names analytics warehouses (Snowflake, BigQuery, Databricks, Amplitude) as the integration target when they are the source of record. Most merchant reads go to the warehouse, not to the platform. Second, every write in Table B is a staged change. Your platform receives it only after a person approves it.

Adobe Commerce teams have the most work in the merchant column, because reporting and campaigns live outside the platform. Our Adobe Commerce developers usually connect the merchant agent to the warehouse first and to MSI second.

The B2B gap: ten changes before a B2B store can use Claude Commerce

The four examples in the blueprint are B2C. ACME sells tents to consumers. The README states that the same interfaces cover other business shapes: on a marketplace, seller becomes a search dimension; with account or contract pricing, the price quoted is the session account’s; with no checkout of your own, the cart hands off to a quote, a purchase order, or a hosted checkout URL. Those are the hooks. The work is yours. Elogic Commerce builds B2B stores, and this is the list we check first.

Two more items sit outside the code. Punchout buyers (cXML or OCI) need the cart to hand off to the punchout return, not to checkout. Large quotes need a sales representative. The agent stages the quote and notifies the representative. Neither is hard. Both must be planned before the pilot, or the pilot measures the wrong thing.

Cost and latency: the numbers to plan with

Claude commerce optimization is mostly caching and model choice. The engineering guide publishes the numbers below. Use them to build the pilot budget.

The request order is the detail most teams get wrong. A request reads from cache up to the first byte that differs from a previous request. Put the global segment first (system prompt and tool definitions, byte-identical across sessions), the session segment second (user context, history, memory), and the volatile segment last (current time, current page). A timestamp at the top of the system prompt breaks the cache on every request. Load skills as tool results, not into the system prompt, so the skill body lands in the cached prefix.

Model choice is a measurement, not an opinion. Anthropic recommends Opus as the starting point for merchant agents, because their tasks are analysis-heavy, and Sonnet for shopping agents, where latency weighs more. Then run the full eval suite across every model and effort level you would consider. Measure cost per completed task, not per model call. A cheaper model that needs more turns is not cheaper. When the result is close, choose intelligence.

How to use Claude Commerce: a 30-day pilot plan

The README recommends that a shopping pilot implements search and product details first and stubs the rest. A stubbed method returns an unavailable result and changes no prompt bytes, so the cache holds. This plan follows that advice.

Week 1. Run the demos and pick the scope

Clone the repository and run the retail vertical. Python 3.11 or later and Node 22 are required.

git clone https://github.com/anthropics/commerce-agents.git && cd commerce-agents
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env        # add ANTHROPIC_API_KEY
(cd examples && npm ci)
python scripts/run_demo.py retail        # API :8000, storefront :3000

Add --merchant to start the portal, or --all for both. Read docs/safety.md (the enforced rules), docs/backends.md (how to map your systems), and docs/deployment.md (Vertex AI, Bedrock, Foundry). Decide one flow to pilot. For most stores, that flow is search plus product details.

Week 2. Implement search and product details, stub the rest

Implement the StorefrontBackend methods for search and product details over your existing search engine. Do not build a new search index for the agent. Leave cart, orders, and policies stubbed. Turn off systems you do not have with the enable_* switches; each switch removes the tools, prompt lines, and grounding rule on every runtime. Use the Claude Code plugin to scaffold the project against your stack.

claude plugin marketplace add anthropics/commerce-agents
claude plugin install commerce-builder@claude-commerce-agents
claude
/scaffold-commerce-agent a shopping assistant for our store

Week 3. Write evals with the people who see the failures

Write 50 to 100 eval cases for the pilot flow with subject-matter experts from customer care, category management, and legal. Construct each case as a snapshot: the system prompt, the tools, and a messages array that puts the agent in the test state. Grade the final state and the rendered response, not the path. For every “should serve” case, write a “should refuse” case. Include injection cases planted in product names and reviews. Then run the model and effort sweep.

Week 4. Add the cart, ship to a canary cohort, measure

Implement the cart methods and the checkout hand-off URL. Roll the agent to a canary cohort. Keep a switch that turns off one skill without a deploy. Measure four numbers: cart size, purchase completion, cost per completed task, and cache hit rate. Compare cart size and completion against the same cohort without the agent, not against Anthropic’s reported figures.

For a merchant pilot, the README suggests a different shape: implement the eight read methods and let the writes refuse. Digests and metrics then run with no write path. Add staged writes after the operators trust the reads.

Claude Commerce readiness checklist

Twelve checks before the pilot starts. A store that passes all twelve can run the four-week plan. A store that fails more than three should fix the platform first. Elogic Commerce scores these items with the Agentic Commerce Readiness Index before a pilot.

1. Ranked search API. Your search engine returns ranked results through an API. The agent must not rank products.
2. Product data quality. Attributes, availability, and prices are complete and consistent. The agent grounds every answer in this data.
3. Cart API that accepts server-issued IDs. The cart must accept only product IDs that the server returned in this session.
4. Checkout hand-off URL. A URL the host can render. For B2B, a quote, purchase order, or requisition list path.
5. Orders and policies through an API. Order status, returns, and policy text are readable by a backend method.
6. Identity in the session. The signed-in customer context, and for B2B the company and role, reach the agent before the first model call.
7. Inventory per source with lead times. The agent can state availability per warehouse.
8. Analytics source of record. A warehouse or reporting API the merchant agent can read.
9. Approval surface for merchant writes. A portal button, a CLI confirmation, or a tool-approval prompt. Shopware gives you a server-side dry-run preview through Admin MCP; other platforms need a preview built in the host.
10. Eval ownership. Each skill and tool has one owner team. A CI eval set exists. The full suite runs nightly.
11. Data handling for memory. A validator on the write path, a retention period, a delete path wired to account deletion, and a per-region switch.
12. Release process. A canary cohort, a per-skill kill switch, and a freeze before peak periods, the same way you freeze other systems.

About Elogic Commerce

Elogic Commerce is a B2B and B2B2C commerce engineering agency founded in 2009. Our 200+ specialists build and run stores on Adobe Commerce, Shopify Plus, BigCommerce, Salesforce Commerce Cloud, commercetools, Shopware, and Medusa.js. We hold ISO 27001, SOC 2 Type II, and ISO 9001 certifications. We are an Adobe Commerce Silver Solution Partner and a Hyvä Bronze Partner. Our AI products include the Agentic Commerce Readiness Index and AI Chat Core. See our AI solutions for ecommerce.

Is your store ready for Claude Commerce?

• We map the blueprint to your platform, list the gaps, and give you a pilot plan in two weeks.
• Request a readiness score through the form below, or read about our agentic commerce services.

Sources

1. Anthropic, “Building commerce agents with Claude”, September 2, 2026.
2. Anthropic, “A guide to the anatomy of effective commerce agents”, September 2, 2026.
3. GitHub, anthropics/commerce-agents: README, docs/safety.md, docs/backends.md, docs/deployment.md.
4. GitHub, Shopify/claude-for-commerce-examples, September 2, 2026.
5. GitHub, agentic-commerce-lab/shopware-claude-commerce (README, docs/shopware-mapping.md, evals/gates.yaml) and the live demo at agentic-commerce-lab.github.io/shopware-claude-commerce, September 2026.
6. Shopware Agentic Commerce Lab, announcement post for the Shopware reference build, September 2026.
7. Shopware documentation, “Agentic Commerce” extension (UCP, beta).
8. Reuters, “Anthropic launches AI agent blueprints for retailers ahead of holiday shopping season”, September 2, 2026.
9. PYMNTS, “Anthropic Debuts Commerce Agent Blueprint With Visa and Mastercard”, September 3, 2026.
10. Forbes, “Why OpenAI’s Checkout Retreat Spells Trouble For Its Commerce Strategy”, March 10, 2026.
11. Ahrefs Keywords Explorer, “claude commerce” matching terms, United States, exported September 6, 2026.