What I Accidentally Built When I Connected Claude to a Product Catalog

In an attempt to create a straightforward price comparison tool, the author inadvertently developed a system that challenged the limits of AI agents' real-world interactions. Despite numerous AI assistants consistently providing inaccurate price information, the author sought an AI capable of accessing genuine prices from actual stores.

The Model Context Protocol (MCP) server enables AI models to invoke external tools, akin to a USB port for AI systems. By connecting Claude to a product catalog encompassing electronics markets in the US and Southeast Asia, the author discovered seven distinct tools that could be accessed: searching for products, obtaining product details, comparison, identifying the best price, discovering deals, listing categories, and ingestion.

The author posed a query to Claude, requesting the cheapest AirPods Pro 2 available in Singapore. Impressed by the AI's ability to search through the catalog, rank results based on price, and present three options from real storefronts along with current prices, the author's project demonstrated the potential of AI agents in the real world. However, the author cautioned that each additional tool added tokens to the context window, potentially overwhelming the system if too many tools were connected.

To optimize the performance of MCP servers, the author recommended keeping tool schemas concise and minimizing the number of tools in a server. By adopting a minimalist approach, the author could avoid the risk of token exhaustion, which could hinder the system's ability to respond to queries efficiently. The author also suggested starting with a smaller set of tools (3 instead of 7) and prioritizing distribution before content development. Ultimately, the key to success lies in measuring what truly matters.