How it stays safe

Part of datakumi Intelligence (intelligence.md), where it is called the assistant.

datakumi ships an MCP server (mdm_mcp/), so any AI assistant that speaks the Model Context Protocol can work with your datasets: Claude Desktop, Claude Code, Cursor, VS Code (Copilot), Windsurf, and local or private models through clients such as Open WebUI, LibreChat or Continue. datakumi doesn't host or call a model; you choose it.

Planned (ai-assistant.md): sensitive fields, a general field setting whose values AI assistants never get, enforced by the server for AI assistant tokens, and guided setup of linking and value rules.


Who it acts as	You. It signs in with an AI assistant token of yours (a personal token is refused): your organisation, your datasets, your fields and record limits, nothing more; sensitive fields are masked.
Changes	Only into your draft. It can submit a change request when you ask, and the dataset's approvers decide as usual. It can't publish, approve or reject.
Read-only	Tick Read only on its token (or on the hosted server's Connect page) and it can answer questions and run SQL, but not change anything.
Audit	Everything it does is in the audit log as you, marked "AI assistant (MCP)". Change requests it submitted show that tag to their approvers.
Data as data	Record values, descriptions and comments are returned as data; the server tells the model never to follow instructions found in them.

The model provider you choose sees what the assistant reads. If a dataset must not reach it, don't give the token access to that dataset; for example, create a member with a role that only covers the datasets you want to share.