Architectural constraints

Two deliberate design choices shape the platform. Both are trade-offs, made consciously, and both drew real user feedback during the pilots.

Stateless gateway: no accounts, no database

Design choice. The system operates without user accounts, session registration or a central database. Users can download individual outputs (tables, charts or text blocks), but whole conversations are not saved or shared server-side.

Trade-off. This eliminates database administration, data retention compliance and user management overhead, keeping the server lightweight and easy to deploy: there is no database to run, back up or secure, and no personal data to look after.

The cost is real too. Missing chat history was the most repeated product complaint in the pilot feedback: testers compare the chat with the tools they use every day, and against that reference the missing history and sharing are felt immediately.

Two things to keep in mind when weighing that complaint. Being a full chat product was never in scope. And the chat gateway is only one client: the MCP server follows the standard protocol, so it can be plugged into existing third-party chat clients that already handle account management, history and sharing. If a future deployment needs those features in this gateway, that is the point at which adding storage would be worth its extra weight.

Structured presentation: tables and charts

Design choice. Tools return structured data (tables and charts) alongside text responses to ensure answers are auditable and verifiable. Users value this a lot; it is what makes an answer checkable against the source.

UI mitigation. The flip side, reported by several pilot testers, is that tables and charts were sometimes repeated or too verbose. Structured outputs are therefore rendered minimised by default in the interface: the full source records stay one click away without flooding the conversation flow.

A better future fix is to detect when the same table or chart is being drawn a second time and simply not repeat it. That deduplication is not built yet; for now, minimising by default is the mitigation.