Anthropic's open-source commerce agents blueprint: what it means for retailers and developers
Anthropic released a forkable commerce agents blueprint on 2 September 2025, with shopper and merchant reference implementations deployable via Claude API, Bedrock, Foundry, or Vertex.
On 2 September 2025, Anthropic published a Commerce Agents blueprint on GitHub, licensed under Apache 2.0. It is a working reference implementation, not a hosted product or a new model. Engineering teams get the prompts, harness code, tool contracts, and safety patterns needed to run two Claude-powered agents: one embedded in a customer-facing storefront, one used by merchant staff in the back office.
The repository is theirs to fork. Anthropic does not accept pull requests to it and there is no SLA on the reference code. What it offers instead is a compression of several months of agent-architecture trial and error into a starting point that a team can have running in days.
What the two agents actually do
The shopping agent sits inside a retailer’s app and handles the full buying journey: it searches and compares catalogue items, builds multi-item carts, answers order status and policy questions, and retains what a customer tells it across a session. Its capabilities are defined as five skills over four system interfaces (catalogue, cart, order, and policy).
The merchant agent supports the people running the business rather than the customers using it. It surfaces sales performance, maintains product listings, flags inventory and price drift, and drafts promotions and campaigns. Every action that would change a live record is staged as a pending change that a person must approve before it goes live. Nothing writes to a system without human sign-off.
That staged-write pattern is one of several safety guardrails built into the blueprint from the start, alongside pricing and availability accuracy controls, cart and refund authority limits, and escalation paths to a human operator.
How it deploys
The blueprint runs on the Claude API, Amazon Bedrock, Microsoft Azure AI Foundry, or Google Cloud Vertex AI, so teams can deploy to whichever cloud infrastructure they already use. The required runtimes are Python 3.11+ and Node 22.
Four verticals come pre-configured with their own ports: retail (3000/3100), travel (3001/3101), telecom (3002/3102), and entertainment (3003/3103). Each has a README with prompts to try on both the storefront and merchant portal surfaces. In the demo environment, no real order is placed, no card is charged, and no live listing is altered. Checkout hands the cart back to the host application to complete, and merchant writes stay staged until approved.
A Claude Code plugin is included. The command /scaffold-commerce-agent asks about the team’s existing stack, plays the architecture plan back for confirmation, and builds the project scaffold. From there, /add-commerce-flow, /author-commerce-evals, and /review-commerce-agent continue the development cycle or pick up from an existing agent codebase.
Where official connectors exist, the blueprint targets them: Snowflake, BigQuery, Databricks, and Amplitude for analytics; Stripe, Square, PayPal, and QuickBooks for finance; Slack, Google Drive, and Gmail for delivery channels.
Launch partners and early results
Anthropic announced the blueprint alongside commitments from Shopify, Visa, Accenture, and Priceline.
Shopify published its own implementation examples covering a storefront shopping agent over Shopify’s Universal Commerce Protocol and Sign in with Shop, and a merchant agent over the Admin API. Priceline built the latest generation of its AI assistant Penny on Claude, using it to navigate flights, hotels, and car hire across complex multi-option trips in a single conversation. Accenture’s position is that the blueprint gives clients a proven starting point to accelerate deployment, combined with Accenture’s existing retail practice expertise.
The early production numbers are notable. Anthropic’s head of product on the Claude platform, Angela Jiang, cited one partner seeing cart sizes up roughly 30 to 35 per cent, with customers about 60 per cent more likely to complete a purchase. Adobe Analytics data cited in Anthropic’s announcement puts AI-driven retail site visits converting at a 60 per cent higher rate than traffic from other sources generally, which gives some context for why retailers are moving quickly on this ahead of the holiday season.
On the developer side, Wix engineers reported having a working commerce agent accepting prompts within fifteen minutes of starting. Zomato’s engineers noted no blockers and said the blueprint’s built-in practices, including tool iteration limits and prompt caching, matched what they had learned independently building their own agent, meaning teams standing up their first Claude agent skip a significant amount of trial and error.
What this means for your team
If you run a retail, travel, telecom, or entertainment platform: this is a practical shortcut, not a vendor commitment. You fork the repository, own everything you build from it, and connect it to your own systems. The four pre-built verticals mean the scaffolding for your specific sector is already there. Given the appetite for conversational commerce tools ahead of peak trading periods, getting a working agent before code freeze is a realistic target rather than an aspiration.
If you are a developer or architect: the Claude Code plugin commands are worth a close look. /scaffold-commerce-agent generating a stack-specific project plan in one step collapses a meaningful amount of early-sprint work. The fact that Zomato’s team found no architectural conflicts with their existing learnings is a reasonable signal that the blueprint reflects genuine production patterns rather than idealised ones.
If you are a merchant rather than an engineer: the merchant agent covers territory that has historically required dedicated operations headcount. Catching price drift, keeping catalogue listings current, and drafting promotions can all run through a single agent that flags changes for your approval rather than acting unilaterally. Sellers growing revenue without growing the team is the framing Anthropic is using, and the staged-write guardrail is what makes that credible rather than alarming.
One thing worth being clear-eyed about: Anthropic is explicit that this is an intelligence layer, not a storefront. They have no interest in owning catalogues, supply chains, or checkout. There is no advertising or paid placement model baked into agents built on this blueprint. The business logic, authorisation rules, and compliance obligations belong entirely to the team deploying it.
The full repository, the primary announcement, and an engineering deep-dive companion post covering agent architecture patterns are all publicly available.