How Aurora handles your case

How a run flows

Aurora Cloud runs in your browser and on our server; there is nothing to install.

1. Your export is read in your browser. When you upload a message export — OurFamilyWizard, TalkingParents, a texting export, or a CSV — it is parsed into a plain message list on your own computer, before anything is sent anywhere.

2. That list reaches our server in memory only. The message list travels over an encrypted connection and is held in memory for the length of the run. It is not written to a database or long-term disk, and it is discarded the moment the run finishes.

3. The reading is done by Claude. To perform the analysis, the message text is sent to Anthropic’s Claude. This is the one place outside our own systems the content goes, and it is how the analysis happens.

4. The result comes back encrypted. The finished brief is assembled into a single self-contained file, locked with a passphrase shown to you once. We discard that passphrase, so from that point the file can be opened only by the person who holds the key.

What Anthropic does with the messages

We use Anthropic’s Claude to do the reading, so the message text is sent to Anthropic as part of each run. Under Anthropic’s commercial terms — not the consumer ones:

• The inputs and outputs are automatically deleted within 30 days.
• The data is not used to train Anthropic’s models.
• In the rare event that Anthropic flags content through its automated safety systems, it may be retained longer (up to about two years) for that review.

What we can, and cannot, see

The short version: no one here sees your case.

• No one at Aurora reads your messages. They pass through our server in memory only, for the length of the run. They are never written to a database or disk, no one on our side reads them, and they are discarded the moment the run ends.
• We cannot open the finished brief. It is locked with a passphrase only you hold, and that passphrase is never stored on our side. The key lives with you alone.
• We keep only account and run details: your email, your access code, a Stripe payment record, and operational metadata (run time, duration, size, cost, export format). None of it includes message contents or the names of people in the case.

What each finished file records

Every finished brief records the version of Aurora that produced it. A run uses several Claude models across its stages; those specific models are kept in Aurora’s internal run record and available on request. So how a given analysis was made stays answerable, down to the version.

Where Aurora is the wrong tool

If a case has material that cannot, under any circumstances, be sent to a third-party model — sealed evidence, certain privileged material — Aurora is not the right fit, because in every mode the messages are read by Claude. There is no local-model option. For the many people with no privilege to protect — someone representing themselves whom an organization has had to turn away — that concern does not arise.

One more option

If you would rather the messages never reach our server at all, Aurora Bench runs on your own computer and calls Claude directly; only license information reaches us. The messages still go to Claude that way, as they do in every mode.