The AI never writes SQL
Most “chat with your data” tools let a language model generate SQL and hope it is right. Datafor works differently:
- The model reads your analysis model’s metadata, not your raw tables.
- It proposes a structured query: which metrics, dimensions, filters, and time ranges.
- Datafor validates the proposal against the model and runs it under the asking user’s permissions, including row-level security.
- The answer comes back with its rows, the query that produced them, and any disclosures.
An AI answer can never see more data than the person asking.
What the AI Agent does
- Multi-step analysis that plans complex questions, asks clarifying questions, and handles follow-ups.
- Chooses the model for you. Since 10.00, the Agent can pick the right analysis model when the user does not.
- Explains changes. “Why did it change?” breakdowns show which dimensions drove a movement.
- Insight components analyze a whole page, a tab, or selected components inside a dashboard.
- Answers in the user’s interface language, with clear result column names and citations.
Bring Datafor into Claude, ChatGPT, and Cursor
Since 10.00, Datafor includes an MCP server. Any MCP client that supports Streamable HTTP can list the models a user can access, explore their metrics and dimensions, and ask questions, all read-only and under that user’s permissions. See how to connect →
Run it your way
- Bring your own LLM. Provider presets, a Verify and enable check, and per-stage model assignment across the agent’s pipeline. API keys are restricted to admins and never exported.
- Operate it. Usage dashboards for questions, success rate, and latency, plus per-user daily quotas with role-based overrides.
- License it. AI users are a separate capacity alongside Creator and Reader seats.