Data and privacy

This page describes what happens to your data when you connect Subtext to an AI assistant. The formal commitments are in the privacy policy; this is the plain-language summary for the connector.

Signing in

• Signing in from Claude or ChatGPT uses OAuth. Subtext's authorisation server (Auth0) asks you to sign in with Google, and matches you to your Subtext account by your verified Google email the first time. After that, the connection is tied to that Google identity.
• Connecting never creates an account. If there's no Subtext account for that Google address, or it hasn't finished onboarding, the connection is refused.
• Your Subtext password or session is never given to the assistant. The assistant gets its own token, which works only for the Subtext MCP server.
• Access keys, for clients configured by file, are shown to you once and stored by Subtext only as an irreversible hash.
• Every request is checked against your live account. If your account is deactivated, every connection stops working on its next request.

What a connection can see

• Only your own graph. Every tool acts on the one account the connection belongs to. There's no way to read someone else's graph through it.

• Only what Subtext has built. The assistant reads the pages, links and summaries in your graph. It can't reach your mailbox, calendar, files or any other connected source through Subtext.

• A connection made by signing in from Claude or ChatGPT can see your whole graph and can write to it, just as you can in the app.

• An access key sees only its scope, which you choose when you create it and can change at any time:

  • Just me: only your name and role.
  • Working context: you, your organisation and your active projects.
  • Full context: everything.

  Anything outside a key's scope is simply not returned. The assistant can't tell "outside scope" from "doesn't exist", which is why it's told to say "that isn't in the context I can see".

What gets written back

Assistants can make three kinds of change, and are told to make them only for durable information, never for passing chatter:

• Submitting new information (edit_page_body). The text goes through the same ingestion pipeline as a document you upload, which merges it into existing pages or creates new ones. Pages it lands on are marked as coming from your AI, so you can tell them apart from your uploads in the app.
• Renaming a page (rename_page), which changes only its display title.
• Linking two existing pages (link_pages).

You can see and edit everything that was written in the Subtext app. An access key set to Read only can't make any of these changes, and only a Full context key with write access can submit new information.

What Subtext stores from a session

• Subtext doesn't receive or store your conversations. Your chat stays with Claude, ChatGPT or whichever client you use. Subtext only sees the tool calls the assistant makes.

• What an assistant writes is stored as part of your graph, like anything else you add.

• A log of tool calls. Subtext records each call:

  • For a connection signed in from Claude or ChatGPT, each tool call is recorded as activity on your account, with the tool name and a short detail such as the search query (up to 280 characters) or the page's slug. Subtext uses this activity to decide what's relevant to you, for example in Pulse.
  • For an access key, each read is recorded against that key with the tool, the query or page, the key's scope and the time. You can see this list under Reads on the key's page.
  • For both, Subtext counts which pages were read, so it can show you which parts of your context are being used.

  These records are kept for as long as your account exists and are deleted with it.

• What the assistant does with what it reads is governed by that assistant's own privacy policy. Once Claude or ChatGPT has read something from Subtext, Subtext has no control over how that provider retains it.

Pausing, narrowing and revoking

From the Connections page, for each access key:

• Pause: the key stops returning any context until you Resume it. The assistant is told that briefing is paused and how you can restore it.
• Scope: switch between Just me, Working context and Full context. It takes effect on the next request.
• Write access: switch between Read only and Read + write.
• Revoke key: deletes the key permanently. Any client still using it is refused from then on.

For a connection you made by signing in from Claude or ChatGPT, remove or disconnect Subtext in that client's connector settings. Deleting your Subtext account ends every connection and deletes your graph; see the privacy policy for what is deleted and when.

Content safety

Your pages are built from your own documents and synced sources, which can contain text written by other people. Subtext tells every connected assistant to treat page content as data, never as instructions, and the brief Subtext writes for assistants wraps page content the same way.

Contact

Questions about privacy or your data: hello@josephmark.com.au. The full privacy policy covers retention, deletion, processors and your rights.