Testing must cover the server's individual features and the complete model-to-tool workflow.
Unit tests should be written for each tool and resource handler. These tests verify the behavior of the function itself without requiring the complete application.
Schema validation tests confirm that the definitions for inputs and outputs are correct. Parameter descriptions, allowed formats, value limits, and return structures should match the actual behavior of the tool.
Integration tests cover the full request-response lifecycle between client and server. They verify that the connection is established, the capability can be invoked, and the result is returned correctly.
End-to-end tests simulate the complete workflow from the model's request to the tool result and final response. These tests are important because a tool can work correctly when called directly while still being selected incorrectly by the model.
Performance tests show how the server behaves under load. They help identify problems that do not appear during a single local request.
Testing should not cover only successful requests. Edge cases, error scenarios, and rate limits should also be tested. A production implementation needs to behave predictably when an input is invalid, a dependency fails, or a limit is reached.
MCP Inspector supports debugging by showing the live capability list and allowing direct calls. A developer can verify a tool in the Inspector before deciding whether a problem is in the server, the client connection, or the model's use of the capability.
The AI Toolkit for Visual Studio Code provides another testing environment. It includes a model catalogue, a playground for changing model parameters such as temperature, maximum tokens, and top-p, and an agent builder for defining an agent's role, behavior, parameters, and tools.
Together, direct inspection, layered tests, and model-driven testing provide evidence that the server works at the protocol, capability, and workflow levels.