MCP server for the LMS (AI access to learner, course and compliance data)
CITI Program's courses are mostly compliance-driven: human subjects research, GCP, RCR, biosafety, IRB training and the like. That makes an MCP (Model Context Protocol) connection to the LMS especially useful, because most of the questions people ask about that data come down to "who has completed what, and who hasn't." What the server would expose Tools: actions the AI can call, like getlearnertranscript, listoverduelearners or createcoursesummary. Resources: read-only context, like course catalogs, curriculum maps, policy documents and completion reports. Prompts: reusable workflows, like "Build an institution's compliance summary" or "Draft a refresher module." Use cases by audience Institutional administrators (customers' compliance and IRB offices) Ask in plain language: "Which PIs on protocol X have expired human subjects training?" or "Show me everyone whose GCP expires in the next 60 days." Generate compliance reports for audits, accreditation or sponsor requirements without building custom exports. Check whether the curriculum assigned to a role (for example, a CRC at a site doing FDA-regulated trials) matches regulatory requirements. Draft and queue reminder messages to learners who haven't completed their training. Learners "What do I still need to finish, and when does my certification expire?" A study companion grounded in the course content they're enrolled in, which explains concepts without giving away quiz answers. Retrieve and share completion certificates. Customer support Resolve common tickets quickly: duplicate accounts, affiliation changes, missing completion records, transfers from a previous institution. Look up a learner's full history across institutions before replying. Content and instructional design Find every module that mentions a regulation when it changes (e.g., "Which modules reference the pre-2018 Common Rule or the old ICH E6(R2) language?"). Draft new modules, quiz items or case studies that match existing course structure and style, then submit them for SME review instead of publishing directly. Analyze item-level quiz data to find questions that are confusing or poorly discriminating. Internal analytics, sales and account management Track usage at each institution, count completion reports and flag accounts at renewal for churn risk. Identify upsell candidates, such as institutions doing clinical research that don't subscribe to GCP. Design considerations Authentication and scoping: Each user sees only what they're allowed to see. An institution admin sees only their own institution, and a learner sees only their own records. Read before write: Start with read-only tools. Add write actions (enrollments, reminders, record edits) later, with human confirmation and audit logging. PII and contracts: Learner records are personal data, and some customers have FERPA, GDPR or contractual constraints. Confirm what our data agreements allow before sending this data to an AI service, and consider de-identified analytics tools as a separate tier. Assessment integrity: Keep answer keys out of anything learner-facing. Auditability: Log every tool call. Customers will ask about this. Proposed first phase A read-only internal server for our support and account teams, with tools for learner lookup, transcripts, institution completion reports and course catalog search. It's low risk, has a clear return on investment, and builds the authentication and logging foundation we'd need before offering anything to customers.
