When sharing an internal app built with ChatGPT Sites, authentication isn't centralized through shared API keys. Instead, it leverages 'bring your own plugin' architecture where each team member brings their own authenticated connectors (e.g., Slack, Notion). Visitors view the same site template, but content dynamically scopes to their identity and team permissions, making it ideal for teams handling sensitive data.
Rather than relying strictly on rigid SaaS dashboards or maintaining permanent tools, rapid generation via Codex and ChatGPT Sites enables disposable micro-software. Users can spin up highly personalized views—such as a tailored Google Calendar view for a specific business trip or a temporary transit navigator—and discard them immediately after use.
Although initially conceived as an easy deployment layer for Codex-generated frontends to bypass local DevOps, ChatGPT Sites has expanded into infrastructure. It incorporates native components like D1 data storage and R2 buckets, and can actually host Model Context Protocol (MCP) plugins, creating a virtuous workflow loop where Sites hosts both the backend plugin architecture and the consuming site.
Instead of building rigid in-app level editors or input forms for community content, developers can distribute an AI 'skill' directly to users. In Kath Korevec's dungeon crawler game, players install a skill that defines structural constraints (room dimensions, controls). Codex then builds custom rooms locally based on open-ended user prompts, enabling scalable, creative community contributions.
Rather than application developers paying massive API bills to run generative, real-time in-game modifications for their users, architectures are shifting toward 'bring your own inference.' End users or players connect their own model compute/subscriptions to modify game worlds or software environments dynamically in real time, shifting cost burdens and opening new interactive capabilities.
Ultra-fast models do more than shave off seconds of operational waiting; they fundamentally change user psychology. When inference delays occur, users get distracted and drop out of flow state. High-speed generation sparks an accelerating loop where each instant output immediately seeds subsequent creative ideas, shifting AI usage from transactional querying to iterative creative exploration.