Model Context Protocol
Your AI of choice, against a warehouse it did not invent.
Most "talk to your data" products hide a prompt behind a chat box. Coeleste exposes a real MCP server. A warehouse guide that teaches the model the schema. A metric registry of definitions you signed off on. Read-only query tools. ChatGPT, Claude, Cursor, and the in-app assistant all speak the same protocol against the same lake the dashboards draw.
The loop
- You ask. Show MRR by product for the last 12 months, and who is about to churn.
- The model reads the warehouse guide, looks at the locked definitions, then runs a read-only query.
- Coeleste runs the SQL. Not the model. The database.
- You hang a tile on that definition. A different window is still that definition. Save it. Revise it. The revision history is still there if you regret the change.
- Tomorrow, an agent asks the same board over MCP and gets the same number.
Why MCP, not a proprietary chat
A proprietary chat dies with the vendor. MCP is how the rest of the AI stack already wants to talk to tools. If your team lives in ChatGPT, Claude, or Cursor, Coeleste is already a tool they can add. If they live in the UI, the same tools run there. One permission model. One warehouse. No second source of truth.
That split matters. The dashboard is where you look. The AI is where you ask. MCP is how the asking stays honest.
The number is never invented
An executive at a reference deployment nearly killed a dashboard program because he believed AI makes the numbers up. He was not wrong about the category. He is wrong about Coeleste. Known KPIs are deterministic SQL. Novel questions go through MCP with an inspectable query. If the query is wrong, you see it.
What you plug in
MCP is only as good as the lake behind it. Coeleste keeps the records and the history. Finance, CRM, meetings, collaboration, the MSP stack. Read-only. We ingest. We never write back.
Coeleste is coming soon. Get early access if you want in. More notes.