DeepEval tutorial: test MCP servers with a local LLM

This tutorial demonstrates MCP server testing with DeepEval through an e-commerce application. The speaker describes an application that uses a locally running language model, vector data stores and a RAG pipeline for product search. A Claude Desktop demonstration accesses the application's local MCP server to find highly rated running shoes, select a size and color, add a shoe to the cart and retrieve the cart contents.

The testing example turns that shopping conversation into a multi-turn test. The speaker explains that the setup needs an MCP server, a configured client and a language model to choose tools. Conversation history and session state matter because later requests depend on earlier results. DeepEval's multi-turn MCP use metric and a conversational test case evaluate the interaction, with pytest running the test and a language model acting as judge.

The speaker reports that both the application model and evaluation run on the local machine, with memory use reaching 69 GB during execution. That figure describes this demonstration, rather than a minimum hardware requirement. The result walkthrough includes product selection and an order summary. This is a test execution overview rather than a complete setup guide: the speaker refers viewers to a separate course for detailed code and STDIO or SSE client configuration.

The local model is used for the test application and evaluation run; the separate Claude Desktop demonstration does not establish offline Claude inference. The presenter promotes a paid course containing the fuller setup and code.