Security and data
Data handling
Nothing leaves until you ask for something. What travels is the records a task needs, at your direction — and what Brella keeps about it is a record of the connection, never a conversation.
What reaches Claude#
Nothing, until you ask for something. There is no sync, no export and no background copy of your book. When you ask Claude to do something, it calls a tool; the tool returns the records that task needs, and those records travel to Claude at your direction.
- A search returns matching policies — the ones you can already see, and no others.
- Opening a policy returns that policy: the insured, the locations and buildings, coverages, documents’ names, the stage. Not the whole book.
- A document is fetched only when a tool asks for it, and PDFs are transcribed to text before they travel.
- A policyholder’s contact details reach Claude only through a tool that needs them — writing to them, or reading the conversation on their policy.
What Brella keeps#
- The connection
- Who connected it, which agency workspace, which app, the permissions granted, and when it was last used. Visible to you in Settings → Claude.
- Token digests
- SHA-256 hashes only. A token cannot be read back out of the database.
- The change itself
- Whatever the tool wrote — a building’s construction, a certificate, a reminder — exactly as if you had typed it into the workspace, on the policy’s activity log with your name and “via Claude”.
- Server logs
- Which action ran, for whom, and whether it succeeded. Not what you asked Claude, and not the contents of the records.
What Brella never has#
- Your conversations with Claude. They are between you and Anthropic. Brella is not in them, is not sent them, and stores none of them.
- Readable tokens. Only one-way hashes are stored, so there is nothing to leak that could be replayed.
- Your Brella OS password. You type it on our sign-in page, in your browser. It never goes near the client.
Third parties, on your behalf#
A few tools use the same outside services Brella OS itself uses when you do the same job on screen. They act on your behalf and under Brella’s own accounts, not yours:
- Address and property lookup.
lookup_addressandlookup_propertysend the address text to our property data provider, which is how an address becomes a real, complete one with a public record behind it rather than something typed from memory. - Reading a PDF. A document filed on a policy is transcribed by our AI provider under Brella’s own account, the same way the workspace transcribes one.
Nothing new is introduced by connecting
Every third party the connector reaches is one Brella OS already uses for that same job. Connecting Claude does not add a processor to your data, and the list in our privacy policy is the whole list.
Anthropic’s side of it#
How Claude handles what reaches it — retention, training, human review — is governed by your own agreement with Anthropic and by their privacy policy (anthropic.com/legal/privacy). Brella makes no promise on Anthropic’s behalf, and you should read theirs as well as ours. If your agency needs a particular arrangement — enterprise terms, no training, a specific region — that is a conversation to have with Anthropic about your Claude plan, not a setting in Brella.
How one agency’s book is kept from another’s#
Tenancy is not a filter applied at the end. Every query in the customer schemas carries the owning agency’s id, the session derives that id from the grant rather than from anything a client sends, and a connection holds exactly one workspace. Brella’s own platform staff are excluded from agency data entirely and cannot connect at all.
The binding statement#
The statement that binds is the privacy policy, section “Connecting Claude to Brella OS”, which carries the same facts as this page plus retention. For anything not answered there, write to brella-os@brellapartners.com — a person reads it.
A data question that is really a permission question — “can a User see this policy?” — is answered on permissions and approval.