The challenge
BoxPlanner had grown for a decade: hundreds of API endpoints, a rich domain model, lots of admin routines that only a few people knew. We wanted box owners and our own support to work with it in natural language ā "which members haven't checked in for 30 days?", "move all Monday classes to room 2", "find duplicate accounts" ā without rebuilding the application.
The solution
We built a Model Context Protocol (MCP) server on top of the existing .NET API:
- 200+ typed tools generated from the domain: members, contracts, schedules, payments, workouts, PRs, SEPA exports, App Insights queries
- authentication and per-box permission checks reused from the platform ā the assistant can only do what the signed-in user may do
- a knowledge base tool so the assistant answers "how do Iā¦" questions from the product documentation
- deployable as a hosted connector for Claude, ChatGPT-style clients and internal agents
Result
Support requests that used to need a developer with database access are now handled in a chat by the box owner. Data clean-ups (duplicate users, exercise remapping, contract regressions) run as conversations instead of scripts. The same pattern works for any .NET application with an API ā that is the service we offer as AI integration.