How does an MCP integration project work?

We first clarify the use case and then check which data and functions are available in the existing system. What the customer provides depends on the interface: data or an accessible application, documentation and contacts are helpful, but there is no universal requirement. The more information and access are available, the more targeted the integration can be. We then define the required MCP tools, develop or extend the interface and describe how the AI should use it. We create several functional tests per interface to check expected results, invalid permissions and erroneous input. These tests are usually written in TypeScript or Python with common test frameworks. Their results can be emitted as logs and integrated into CI/CD depending on the customer project; not every customer wants a CI/CD pipeline. Without a pipeline, developers or AI can run the tests, including as a recurring task such as a weekly check. We recommend blocking failures involving permissions, data access or write and delete actions. Non-critical failures in read functions should at least produce a clear warning and be assessed; the customer decides who may approve a release despite a warning. Finally, we test the integration together with the customer and agree which actions and confirmations are required.

MCP integration in a project