The value over pointing a model at the warehouse is the semantic layer: definitions, row-level and object-level security already live in ThoughtSpot, and the answers arrive through them. It also writes back — Liveboards and embeddable visualizations get created from the same conversation that produced the question.
ThoughtSpot's Spotter server, hosted by ThoughtSpot: it authenticates against your instance and answers from your own data through the semantic model already defined there.
- Natural-language questions answered from live data in your ThoughtSpot instance, not from a copy
- Liveboards and visualizations created programmatically out of the conversation
- Embeddable visualizations generated for use elsewhere
- Search across existing ThoughtSpot content by name, author, tag or type
- Answers that respect the governed semantic layer, including row-level and object-level security, so no analytics logic has to be rebuilt on your side
- Date-based API versioning: pin the tool contract by appending an api-version to the endpoint, or leave it off to track the latest
A ThoughtSpot instance and an account on it. The hosted endpoint is https://agent.thoughtspot.app/mcp over streamable-http, with SSE alongside it. Clients that cannot do dynamic client registration register manually at https://agent.thoughtspot.app/clients — the OAuth client ID and secret are shown once and never again — and that registration can pin a ThoughtSpot URL so users are not asked for it. There is also a bearer-token endpoint that takes an Authorization header, with the instance either folded into it as token@ts-host or sent separately as X-TS-Host. Published on npm as @thoughtspot/mcp-server for running it yourself.
One command — npx mcp-remote https://agent.thoughtspot.app/mcp?api-version=latest