The Model Context Protocol (MCP) is getting a lot of attention, both positive and negative. This talk is about what happens when a team tries to ship it as a production service to approximately 300 enterprise tenants, each with their own authentication, their own data, their own ways of using the platform, and their own AI client of choice. At Thrive Learning, Alex Tasioulis and his team built and deployed an MCP server that exposes their learning management system/learning experience platform (LMS/LXP) to AI agents. The engineering challenges were significant: how can users be authenticated across hundreds of separate Amazon Web Services (AWS) Cognito user pools without a per-tenant deployment? How can sessions survive rolling AWS Elastic Container Service (ECS) deployments? When tenants span healthcare, retail, and aviation, and each uses the platform differently, how can a team design a tool surface that is genuinely useful across all of them without building bespoke integrations for every use case? How can teams reconcile usage patterns of both end users navigating their own learning and admins managing content and people at scale? And once the system has been built, how can it be tested to ensure the workflows work as expected? When things go wrong, how can teams debug issues when they can never see the original prompt the user typed? This talk walks through the real architecture Alex and his team built, with URL-path tenanting and live tenant resolution via an internal discovery application programming interface (API), pre-registered app clients without dynamic client registration (DCR), and discusses practical lessons from testing across Claude, Microsoft Copilot, Gemini, and more, as well as their approach to MCP observability, evaluations, and what they wish they had built earlier. The time of demos is over. This is what it looks like to engineer MCP for a real business-to-business software-as-a-service (B2B SaaS) product.