anthropics/commerce-agents

What it solves

This project provides a reference implementation for building AI-driven commerce experiences. It addresses the challenge of creating specialized agents that can handle both customer-facing shopping tasks and internal business operations (merchant management) while maintaining safety, guardrails, and integration with existing business systems.

How it works

The system consists of two primary agent roles:

• Shopping Agent: A customer-facing assistant that handles product search, comparisons, planning, cart management, and policy inquiries. It uses a StorefrontBackend to interface with a company's catalog and order systems.
• Merchant Agent: A back-office tool for staff to analyze performance, update listings, manage inventory, and draft marketing campaigns. To ensure safety, all "write" actions (changes) are staged and require human approval before being applied.

These agents can be deployed in three ways: via the Messages API (for custom host applications), the Agent SDK (for SDK-managed loops), or as Managed Agents (using MCP servers).

Who it’s for

• Developers looking to build AI shopping assistants or merchant portals.
• Businesses wanting to implement LLM-based commerce automation with built-in safety gates.
• Architects designing the integration between LLMs and commerce backends (catalogs, carts, and analytics).

Highlights

• Dual-Role Architecture: Separate specialized agents for customers and merchants.
• Safety First: Includes fencing, provenance gates, and a mandatory human-approval loop for merchant changes.
• Flexible Deployment: Supports multiple runtimes including the Messages API, Agent SDK, and Managed Agents.
• Industry Verticals: Includes four runnable examples covering retail, travel, telecom, and entertainment.
• Scaffolding Tool: Includes a Claude Code plugin to help developers scaffold new commerce agents based on their own tech stack.