Building commerce agents with Claude

Overview

Anthropic released a blueprint for building commerce agents on Claude on September 2, 2026 — a working reference implementation, so that catalog search, carts, checkout handoff, and approval flows don’t have to be designed from scratch for every commerce agent. Retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase. The code is available today under Apache License 2.0 at anthropics/commerce-agents.

Key features

• A shopping agent and a merchant agent

  The blueprint ships two agents. The shopping agent searches catalogs, assembles multi-item requests, personalizes recommendations from customer preferences, displays products and comparisons in-conversation, builds carts and hands off to checkout, and answers customer service questions about orders and returns. The merchant agent covers the store side: analytics on sales performance, inventory tracking with alerts, pricing and promotion recommendations, and drafting marketing campaigns.

• Two agents defined once, deployable three ways

  The agent definitions — types, backend interface, prompt, tools, and gates — sit apart from the runtime, so the same definition runs on a Messages API loop, on the Claude Agent SDK, or on Claude Managed Agents. Rolling your own usually welds the prototype’s execution path into the code and makes moving later expensive; here you swap the runtime directory instead. Model access works through the Claude API, Amazon Bedrock, Microsoft Foundry, and Google Cloud Vertex AI.

• Four runnable verticals

  Retail, travel, telecom, and entertainment each ship as a pair: a storefront the customer sees and a portal the operator sees. Each vertical carries its own hard parts. Travel handles date-bound inventory and itinerary extension; telecom handles account context, a plan matrix, and regulated fee disclosures; entertainment handles timed holds, waitlists, transfers, a venue map, and all-in fees. Each example README has a Try section with test prompts.

• Scaffold your own with the Claude Code plugin

  Rather than editing an example down to fit, the commerce-builder plugin scaffolds a custom agent. Installing it takes two steps:

  claude plugin marketplace add anthropics/commerce-agents
  claude plugin install commerce-builder@claude-commerce-agents

  The plugin adds /scaffold-commerce-agent to generate a new agent, /add-commerce-flow to extend an existing one, /author-commerce-evals to create evaluations, and /review-commerce-agent to review an implementation.

• Writes go through an approval gate

  When the merchant agent restocks inventory, fixes a listing, or moves a price, nothing applies immediately: the write lands as a staged change and an operator approves it first. Alongside that, fencing and provenance gates run on tool calls across all paths, memory has validation and caps, and analysis runs under budgets. The rules are documented in docs/safety.md.

• Named integration points

  The customization guide names what to change when wiring the blueprint to a real system. Backend methods call your services server-side and the model only ever sees results. The backend enforces step sequences for fixed-order flows, checkout hands off to your own route or a platform-hosted URL, product options stay flexible, and cart, inventory, pricing, and campaign features toggle individually through config switches.

Where to go next

The announcement points to three resources in one sentence: “It’s available today, with live demos for each vertical and an engineering deep-dive on how it was built.”

• The repository — what “available today” links to. The whole blueprint under Apache 2.0 in anthropics/commerce-agents: both agents’ tools, prompts and skill definitions, the four vertical examples, the commerce-builder plugin, and the safety rules in docs/safety.md.
• Live demos for each vertical — the “live demos” link. Try the retail, travel, telecom, and entertainment examples in the browser before cloning the repo and setting up a local environment.
• The engineering deep-dive — “A guide to the anatomy of effective commerce agents,” published the same day. It covers why the architecture uses a single model with skills rather than subagents, how prompt caching cuts latency and cost, and how memory, safety enforcement, and evals are run in production. Useful for the design reasoning even if you never fork the blueprint.

Notes

• Prerequisites — Python 3.11+ and Node 22. Clone the repo, install dependencies, add ANTHROPIC_API_KEY to .env, and run python scripts/run_demo.py retail to bring up the API on port 8000 and the storefront on port 3000. Add --merchant for the portal instead, or --all for both.
• All data is fictional — the examples run on data from a made-up company, ACME, with no live orders or charges.
• No MCP connectors ship — integrate your own, or use official connectors such as Snowflake, BigQuery, Stripe, PayPal, Slack, and Gmail.
• It is a reference implementation — published under Apache License 2.0, but not maintained and not accepting contributions.
• Partner routes — Anthropic named Accenture, Mastercard, and Visa as solutions and ecosystem partners working with them to enable clients and merchant communities to leverage the blueprints. The announcement cites Shopify and Priceline among enterprise customers already running consumer-facing commerce agents on Claude.
• Where the 35% and 60% figures come from — both are reported by retailers already running shopping agents on Claude, not a projection of what adopting this blueprint will produce.