Lectures

We break down why directly connecting ChatGPT to a single MCP server quickly runs into a “God object” and a URL zoo, and why an MCP Gateway shows up in prod: a single entry point, tool routing ( tools/list, tools/call), token verification, observability, rate limiting, context enrichment ( openai/locale, _meta["openai/userLocation"]), and advanced routing (locale, canary, service versions).

We break down why the ChatGPT Apps “perimeter” is easy to overheat: the LLM can invoke tools repeatedly and in parallel on its own. We assemble a practical protection scheme: a proxy/gateway in front of MCP, meaningful rate limiting (by userId/ tenantId), offloading heavy operations into a queue with 202 and jobId, plus backpressure (limiting active calls and queue size). Separately, we discuss webhooks, retries, and “storms,” plus a mini‑practicum and a list of common pitfalls.

In this lecture, we break down practical “resilience” for ChatGPT Apps and MCP/Gateway: why stuck tool-calls break UX and degrade the model’s answer quality, how to set timeouts correctly, apply a circuit breaker (with closed/ open/ half-open states), isolate resources via bulkheads, and how to survive webhook storms (duplicates, retries, spikes) with a queue-first approach and idempotency. With TypeScript examples we show fetchWithTimeout, mini-implementations of CircuitBreaker and Bulkhead, and a simplified webhook handler in Next.js.

This lecture is about how to turn “one instance on Vercel” into a resilient production system for ChatGPT Apps: horizontal scaling and a stateless design for the MCP Gateway and backends, L7 load balancing with health checks and SSE support, splitting infrastructure into clusters by workload type (gifts / agents / commerce), as well as safe release strategies ( blue/green, canary) and fast rollback with versioning and diagnostic endpoints.

In this lecture we break down how to speed up ChatGPT Apps using HTTP caching and the edge layer: what exactly CDN/ edge cache give you, how to set Cache-Control correctly (including s-maxage and stale-while-revalidate), when ETag and a 304 response are useful, how to combine SWR at the CDN level and at the UI level, and why Vercel Edge Functions are not a “mini‑backend” but a tool for light logic and routing with strict limits.