Scope

Decide exactly what an AI can see

This is the heart of it. You hand intelligence a slice of your knowledge — a whole source, or a bounded project scoped by rules you set — and nothing outside that line is in the room. The boundary is enforced where the data is read, not promised on a settings page.

AvailableAccess
default-denyscope lock

Your sources, your switches

Work notesobsidian
Visible to AI
ChatGPT historychatgpt
Visible to AI
Personal journalday one
Hidden
Client contractsfilesystem
Hidden
Readwise highlightsreadwise
Visible to AI
When the assistant searches, hidden sources simply don't exist to it — no results, no snippets, no citations.
Archiving something is not the same as exposing it — you decide, per source, what any AI may read.
01

Hand an AI your whole memory and you've traded one risk for another. Useful context has to be granted, scoped, and revocable — not siphoned wholesale.

02
  • Grant access at the grain you want: a whole source, or a bounded project scoped to exactly the sources you choose.
  • A per-source switch turns any source off-limits to every assistant, instantly.
  • Every actor — you, an assistant, an importer — runs with an explicit identity and scope; default-deny when no grant applies.
  • Scope can tighten but never silently widen mid-task; the boundary is enforced in the query layer.
03
  1. Choose what an assistant may see — a source or a bounded project.
  2. It runs against that slice and nothing else.
  3. Revoke or narrow the grant whenever you want.
  4. Check what any actor was allowed to reach.
04
  • Nothing runs without a known actor and scope, and no action can quietly widen its own access.
  • Even you act through a scoped identity — there's no master key that silently sees everything.
  • Today scope runs through sources and projects; one unified grant-anything control — hand-picked sets included — is on the roadmap.