Lectures

In this lecture, we examine the “brain” of ChatGPT Apps: how the model chooses between a regular text reply and a tool call ( tool-call), what exactly it sees (system prompt, the list of tools, the conversation history), why name/ description/ inputSchema are critical, what a call looks like at the API level ( tool_calls, finish_reason), and how permissions/annotations ( readOnly, destructive) influence model behavior and UX.

In this lecture, we cover how to describe tools in MCP and ChatGPT Apps using JSON Schema: how inputSchema is structured, why it is part of the model’s “prompt,” how to write a good description, when to use enum and constraints ( minItems, format). Then we build a “single source of truth” in TypeScript via Zod + zod-to-json-schema, discuss issues with optional fields, and show the role of _meta/annotations (for example, readOnlyHint, destructiveHint, openai/outputTemplate) for UX and safety.

We walk through the server path of tool invocations in the Apps SDK + MCP ecosystem: how ChatGPT calls a tool via JSON-RPC, where the MCP server lives in the Next.js template ( app/mcp/route.ts), how to design the inputSchema and return the result via content and structuredContent, how to bind the UI through _meta and openai/outputTemplate, as well as how to do validation, authorization, work with external APIs, logging, and error handling.

We break down the full data path from the server tool to the React UI in ChatGPT Apps: how ToolOutput arrives in the widget via the Apps SDK, how structuredContent, content, and _meta differ, how to safely read toolOutput via hooks ( useWidgetProps) instead of directly accessing window, and how to build UI states (loading/empty/error) around the tool call result.

In this lecture, we discuss why in ChatGPT Apps the same tool can be invoked more than once (retries, Regenerate, network glitches), how to design idempotent operations via idempotency_key, how to return errors so that both the model and the UI understand them consistently ( ok/ error, code, retryable), and how to safely configure a widget via text/html+skybridge and _meta (CSP, domain, visual flags).