The problem
Useful information sat in a business system that colleagues used every day. Using it with an AI assistant meant copying information between tools and manually applying suggested changes.
What I contributed
I built and extended a connector using the Model Context Protocol, building on existing open-source work. I defined the requirements and access model, directed the AI-assisted implementation, reviewed and tested the changes, and took responsibility for deployment and maintenance.
What changed
Colleagues can search and read information through an AI assistant and, where enabled and permitted, update records. The connector is designed to use the signed-in person's access. Changes are enabled selectively, and usage records help identify requests that need better handling. An internal guide supports colleagues using the connector.
What this illustrates
Making an integration useful means considering access, understandable errors and ongoing maintenance alongside the connection itself.
Technical detailHow the integration works
MCP exposes a business-system API as tools an AI client can call. The implementation uses identity-aware access, separates data between users and enables write tools selectively. Consistent responses and structured refusals help the client understand why a request cannot proceed. Automated tests, deployment checks and operational records support maintenance. The project extends an open-source starting point; it is not an implementation authored entirely from scratch.