Connect Claude or ChatGPT to only the folders you choose

Connecting Claude or ChatGPT to Hjarni used to be all or nothing. Sign in, and the assistant could read every note you have.

That's fine for your own chat assistant. It's not fine when you connect a client's ChatGPT workspace, or a work Claude account that shouldn't see your journal.

Now the sign-in screen asks what the assistant can reach.

Everything, as before. It's preselected. Click Allow as always.

Only folders you choose. Tick a client folder, a project or a whole team. A folder covers everything inside it, including what you add later. The assistant sees nothing outside them, not even their names.

Read only. One switch, for either choice. The assistant can search and read, but not create, change or delete anything. Pair it with everything and it can find anything you wrote and touch none of it.

Change it later. Settings › Connections now says what each assistant reaches. Change access opens the same page as agent tokens: a level per folder, folders set to no access, and instructions for that one assistant.

Search, lists, files and history all stop at the folders you picked, over MCP and the REST API. And it never gets more than you have: in a team where you're a viewer, it reads.

One thing to know: in ChatGPT, the note panel and @-mentions aren't available on a limited connection yet.

Assistants you already connected keep full access until you limit them. New accounts see the sign-in screen as before: until your first note there is nothing to limit.

Live now, on every plan. The details are in Agent tokens.