Your User Isn't Human: Testing MCP Servers

All tests are green, the coverage percentage is high, and the agent responds to the user something like, "I don't have such a tool." The tool does exist. It's covered by tests. The bug isn't located in any line of code. 

This is the biggest surprise MCP servers offer QA: their primary user isn't a human or another service, but an AI agent. The model doesn't "connect to the API"—it reads. All it sees about your server is text: the tool name, description, and parameter schema. This means the contract now consists not only of the schema but also of text, and changing one word in the description changes the system's behavior in production.

In this report, we'll look at how an agent selects a tool and where the scenario breaks down: the agent didn't select your tool, called it with the wrong arguments, or misunderstood the response.