MCP apps fail without data-model separation — Indeed's practical lesson

MCP apps let companies inject branded UI into Claude and ChatGPT, but naive implementations create a black-box problem for the model. Dustin Mihalik from Indeed shares that simply calling existing APIs and displaying HTML leaves the model blind to what the user sees.

The core issue is follow-up capability: when a user clicks 'view details' or asks a question, the model has no context about the displayed data. Mihalik's first rule is that anything shown to the user must also be provided as structured data to the model, using MCP's resource URIs and tool descriptions.

The most critical insight is separating data processing from UI rendering. For complex multi-step tasks like job searches across cities, batch tool calls must remain text-based to let the model iterate freely. Only the final curated results should be rendered as UI components via a dedicated render tool.

This approach creates composable, small tools that give the model flexibility to explore data without losing context. The takeaway for teams building AI interfaces: prioritize data flow and model context over visual polish. UI is a rendering side effect, not the main challenge.