Secure AI Course Access: A Production MCP Server with Entitlement-Scoped Retrieval

The client wanted an AI assistant that answers from paid course material — but only for the courses a given student had actually bought. Two failure modes would each break the business: a student asks about content they never paid for and the assistant answers anyway, or someone uses the assistant to bulk-extract the entire library. Complicating it further, there was no existing auth or purchase API to plug into.

We built a remote MCP server exposing course-scoped retrieval to an AI client. OAuth 2.1 with PKCE and JWKS was designed and built from scratch, since no auth layer existed. Retrieval is filtered by entitlement server-side rather than in the prompt, so scope cannot be talked around. Anti-bulk-export guardrails and audit logging keep the assistant answering questions rather than surrendering the library. We then profiled the endpoint and parallelised the independent reads — rate limiting, entitlements, embedding, citation and audit — to cut response time.