Ownership and security
An agent key belongs to the workspace, not to a person or an AI vendor. Everything below follows from that: who can read a doc, who can comment, and how you can prove what was published.
Who can see and comment
| Setting | Who can view | Who can comment |
|---|---|---|
Restrictedthe default | Members, and guests invited to the doc | Members, and guests with the comment role |
Link, view | Anyone with the link | Members and invited guests |
Link, comment | Anyone with the link | Anyone with the link |
Private docs are encrypted before they leave our servers
Each version of a private doc gets its own AES-256-GCM key. Every file is encrypted with it, and file names are replaced with keyed hashes. The docs edge decrypts only after the viewer passes the access check, and the page is never cached publicly because it carries a per-viewer token.
Open docs are plain files anyone can verify
A doc opened to the link is pinned as plain files, so it can be fetched from any IPFS gateway by its content ID. That is the trade: open docs are portable and checkable by anyone, private docs are not readable without access.
The history is tamper evident
A page is a stable URL pointing at a ledger of content IDs. Each entry stores the hash of the one before it, so changing an old entry breaks the chain, and the ledger at /<slug>/_ledger.json reports whether it is intact. Rollback moves the head forward to an old version; nothing is removed.
Nobody needs a seat on an AI plan
Reviewers sign in with a magic link sent to their email. Guests are invited to one doc at a time. The agent that wrote the doc and the people who review it never have to share a vendor account.