Connect it once.
One entry in your client's plugin directory. The same key as the API, the same rights gating, and read only tools you approve up front.


Ask for the latest Swiss unemployment rate and you get a plausible figure from training data, months out of date, with no way to tell.
The server gives the model a dataset to query. Every value comes back with its source, its period and its freshness, so you can check the answer.
One entry in your client's plugin directory. The same key as the API, the same rights gating, and read only tools you approve up front.

No indicator IDs, no country codes. Ask the way you would ask an analyst and the server resolves the entities and picks the series.

Every value comes back with the institution that published it, the period it covers and whether it is still fresh. You verify the answer instead of trusting it.

All of it comes down to whether you can check the answer.
We run the server ourselves, with the same auth and rights gating as the API.
Every number comes from the institution that published it.
The agent knows when a value has gone stale against its own cadence, and says so.
Only data cleared for redistribution is served, enforced in the query layer.
The same key works for the API and the server.
One block of JSON in Claude, Cursor or any compatible client. No middleware.
Start with a number you already know. Every value carries its source and period, so you verify rather than trust.
One block of JSON in your client config. The same key as the API.
Claude, Cursor, and any MCP compatible client.

{ "mcpServers": { "quantora": { "url": "https://mcp.quantoraresearch.com", "headers": { "Authorization": "Bearer YOUR_API_KEY" } } } }
An open standard that lets AI assistants call external tools. Ask a model about Turkish inflation and it produces a plausible number from training data. With the server connected it fetches the actual print, with the source attached.
What inflation is doing in Turkey. How G10 policy rates compare. US companies growing revenue over 30 percent. What the 10 year and 2 year spread is saying. It resolves the entities, picks the indicators and shows its sources.
It cannot invent a number. It returns what is in the database, or it explains why there is none and what would work instead. An empty result with no reason is the easiest way for an AI client to produce a confident wrong answer.
Yes. Same store, same rights rules. A different doorway, not a reduced dataset.
No. Turkey, TUR and the internal entity id all work. Removing that is the point of MCP.
The MCP server is in private testing while we finish plans and documentation.