AI you can audit — The rest of the trust surface, each one checkable.

Permissions hold through AI The AI runs under your permissions. A field hidden from you never enters AI context — not in answers, not in summaries, not in search. The Ask page lists exactly which fields are hidden from the AI for you, by name, so the guarantee is readable, not asserted. An outside agent is still a person RowFold is an MCP server, so Claude Code, Claude Desktop or Cursor can work your workspace directly. A connection is bound to one member and sees exactly what that person sees: hidden fields stay hidden, no-access tables stay invisible, and there is no separate integration-permissions model to configure or get wrong. Writes are off until you allow them, and land in the audit log as MCP · connection (as person). AI changes are signed Everything the AI creates or changes appears in the audit log as its own actor — AI · Ask (for you) — exactly like API · token and Automation · name. "Who changed this?" has one consistent answer, and an AI write is never dressed up as a human one.