Lectures

In this lecture, we’ll break down what MCP ( Model Context Protocol) is, why the protocol is needed in the ChatGPT Apps stack, and how it differs from “your own REST.” We’ll align on the architecture picture (ChatGPT → MCP client → your MCP server → DB/API), study the key protocol entities ( tools, resources, prompts, and capabilities), and also understand how MCP is already “hidden” inside the Apps SDK through calls like callTool—and how to stop treating it as a black box.

In this lecture, we break down the “raw” MCP message format on top of JSON‑RPC 2.0: how request, reply, and notification differ; how to correlate messages by id; when the server should return result vs. error; and what the primitives tools/ resources/ prompts look like in the protocol (methods tools/list, tools/call, resources/read, prompts/get). We’ll tie this to the Apps SDK using window.openai.callTool and practice reading/fixing MCP JSON.

We unpack the client–MCP server “introduction” via the initialize request: how protocolVersion is negotiated, which capabilities the client and server declare, why it is important to wait for notifications/initialized, and how discovery works after the handshake via tools/list, resources/list, and prompts/list. We finish with practical examples for GiftGen and a set of common pitfalls.

In this lecture we assemble the first working MCP server on Node.js/TypeScript: we bring up an HTTP endpoint /mcp, create a McpServer from @modelcontextprotocol/sdk, attach a transport like StreamableHttpServerTransport, and register three basic MCP entities: a tool (the suggest_gifts tool with a Zod schema inputSchema), a resource (the gift_catalog directory with a proper mimeType), and a prompt (the birthday_gift template). Plus—we’ll cover the runtime, local launch, and common pitfalls.

In this lecture we will break down practical debugging of MCP servers using inspectors: how MCP Jam/Inspector repeat the handshake (the initialize method) and read capabilities, how to manually hit tools/list and tools/call and view raw JSON-RPC, and how to set up logging on the server (structured JSON logs, separating stdout/ stderr, PII masking) so that issues are caught before integrating with ChatGPT Apps / Apps SDK.