Detailed Analysis
Tredict, an endurance sports training platform built since 2020, became one of the early third-party applications listed in OpenAI's ChatGPT App Directory following the directory's opening to external submissions in December. The developer behind the platform documented a striking asymmetry in the build process: two weeks of active programming followed by three months of review by OpenAI before the app went live. The application allows users to connect their Tredict accounts to ChatGPT and perform natural-language actions — analyzing past activities, renaming sessions, and generating structured workout plans — with planned workouts syncing to major sports hardware including Garmin, Coros, Wahoo, and Suunto devices.
One of the more technically distinctive aspects of the Tredict integration is its use of MCP UI Apps, a pattern that renders interactive widgets — complete with charts, maps, and activity metrics — directly inside the chat thread as sandboxed iframes. The developer notes this widget-based approach remains uncommon among ChatGPT integrations, most of which use tools only. The central engineering challenge was delivering user-authenticated content into those iframes, which have no access to OAuth tokens by design and for which no documented best practices yet exist. This represents a frontier problem in the emerging ecosystem of AI-native application development, where the conventions for embedding rich, authenticated UI within conversational interfaces are still being established.
The post offers a candid real-world comparison between Claude and ChatGPT as MCP host environments. The developer observes that ChatGPT is conservative with its context window usage, tending to fetch surface-level activity lists rather than pulling full metrics unless explicitly instructed, which produces shallow responses to vague prompts. Claude.ai, by contrast, is described as performing "noticeably better" for complex, multi-week training plan creation — capable of generating periodized plans spanning weeks or months, incorporating mixed sport types and individualized intervals derived from historical activity data, a task with which ChatGPT reportedly struggles. Critically, the developer confirms that the same MCP UI App widgets — the interactive activity and plan views — render correctly within Claude.ai as well, meaning the integration is genuinely host-agnostic.
The situation creates a notable irony that the developer explicitly calls out: the platform works best with Claude.ai, yet Tredict's application to be listed in Anthropic's connector directory has been pending without any feedback or rejection rationale. This stands in contrast to OpenAI's process, which, while slow at three months, ultimately produced a result. The public appeal for a status update from "any Anthropic folks" who might see the post reflects a broader pattern in the AI developer ecosystem, where access to distribution channels — app directories, connector listings, and curated integrations — is becoming as strategically important as technical capability. Developers building on MCP infrastructure are now navigating multiple bureaucratic pipelines simultaneously, with outcome timelines that bear little relationship to development effort.
The Tredict case illustrates the maturation and growing complexity of the AI application layer. The Model Context Protocol, which allows a single server to serve multiple AI hosts, is enabling developers to build once and deploy across competing platforms — but the gatekeeping mechanisms at each platform remain inconsistent and opaque. As MCP adoption grows across Anthropic, OpenAI, and other AI providers, the policies governing which third-party applications gain visibility within those ecosystems will increasingly shape which use cases get built, which niches get served, and which developers can build sustainable businesses on top of frontier AI infrastructure.
Read original article →