Anonymised. Figures rounded.
Case study · A market-intelligence data company
An MCP server for a market-intelligence data product
A data company wanted its curated datasets usable from AI assistants without handing an agent raw SQL. The server guides the agent: list the datasets, pick a view, learn the relationships, then filter, aggregate or profile, with variants that return ready-made UI.
The figures
Ranges and rounded numbers, added as they are confirmed. None are estimated.
- Datasets exposed
- 5+
- Tools exposed
- 15+
What was built
- A dataset catalogue tool, so the agent starts from what exists.
- View selection, so every question runs against a known shape.
- Relationship hints, so the agent never guesses a join.
- Filterable-attribute discovery, so filters use real values.
- Filter, aggregate, profile and summarise tools.
- UI-returning variants for the answers a person reads.
Why it mattered
The agent stops guessing. Answers are reproducible, because every one runs through the same steps. And the company can see what people ask.