An MCP server for product catalogs should function as a semantic wrapper around your existing data, not a thin API proxy. The goal is to make product data agent-ready: understandable and actionable by AI models during real-time conversations.
Recommended Tool Structure
Organize your catalog tools into three categories:
Discovery Tools
These help AI agents find and explore products:
- catalog_search_products: Natural language product search with filters for category, brand, price range, and availability. This is the most-invoked tool in most deployments.
- catalog_list_categories: Browse your product taxonomy and category hierarchy. Agents use this to narrow searches and understand your catalog structure.
- catalog_get_product_details: Retrieve full product information by ID or SKU, including pricing, specifications, reviews, and availability.
Recommendation Tools
These enable personalized product suggestions:
- catalog_get_recommendations: Personalized recommendations based on context, user behavior, or conversation history.
- catalog_get_similar_products: Find similar products using vector similarity or attribute matching.
- catalog_get_trending: Surface trending, popular, or seasonal items.
Operational Tools
These enable transactions within AI conversations:
- cart_add_item: Add products to a cart with quantity and variant selection.
- cart_get_summary: Retrieve current cart state, totals, and applied discounts.
- order_get_status: Check order status, shipping, and delivery estimates.
Exposing Context via Resources
In addition to tools, expose structural context as MCP Resources that the AI model can read:
- catalog://schema: Your product data dictionary defining available fields, types, and relationships.
- catalog://taxonomy: The full category hierarchy so agents understand how products are organized.
- catalog://pricing-rules: Pricing model documentation including tiers, discounts, and currency handling.
Resources give the AI model context without requiring a tool call, reducing latency and improving response accuracy.
Prompts for Common Workflows
Define MCP Prompts that provide few-shot examples for frequent agent tasks:
- product-search-assistant: Examples showing how to interpret natural language queries and map them to search filters.
- recommendation-explainer: Templates for explaining why a product was recommended, including comparison points and tradeoffs.
Schema Design Principles
When designing your catalog schema, follow these principles:
- Expose taxonomy: Share category hierarchies, affinity clusters, and co-purchase patterns so agents understand product relationships.
- Surface behavioral signals: Let recommendation engines expose similarity metrics and personalization weights.
- Use structured outputs: Return product data in consistent, typed JSON so agents can reliably parse and compare results.
- Implement vector search: Use embedding-based search for natural language product queries that go beyond keyword matching.
- Batch operations: Use the Tasks primitive for bulk catalog operations like price updates, inventory syncs, and seasonal promotions.
Testing Your Catalog
Use the MCP Inspector to verify:
- Search tools return relevant results for natural language queries.
- Product details include all fields agents need for comparison and recommendation.
- Resources load correctly and provide useful context.
- Error handling returns actionable messages agents can interpret.
Sources
- MCP Official Documentation: https://modelcontextprotocol.io
- MCP Specification (Tools): https://modelcontextprotocol.io/specification/2025-11-25
- commercetools Commerce MCP Reference: https://commercetools.com/commerce-platform/commerce-mcp