Privacy
Publishing a page publishes the work. It should not publish you. This page says exactly what we send, what we strip automatically, and — the part most services leave out — what we cannot strip reliably and are asking you to check yourself.
What a published page contains
- The title and prompt you approve in the preview.
- The result itself: the model, splat or HTML bundle, and its card image.
- A receipt: one line per agent action, holding the tool's name, the first 200 characters of its arguments after scrubbing, a timestamp, and token counts.
- Totals derived from those lines: tool calls, turns, tokens, elapsed time.
What never leaves your machine
- The raw agent session file, in full.
- Every tool's output — file contents, command results, screenshots. Only the arguments are summarised, never what came back.
- Your files, environment variables, shell history and credentials.
- Anything in your working folder that is not part of the result you chose.
What we remove automatically
The scrubber runs on your machine before anything uploads, and again on our
server before the receipt is stored. It replaces each of these with a marker such
as [EMAIL], so you can see in the preview that something was taken
out:
Secrets
- Private keys and SSH public keys
- API keys and provider tokens
- Passwords inside URLs
Authorization,Cookieand API-key headersKEY=valuepairs whose name looks like a secret
Contact details
- Email addresses
- Phone numbers
- Payment card numbers and US social security numbers
- GPS coordinate pairs
Your machine
- Home directory paths, rewritten to
~ - Your local username
- Temporary folder ids
- Hostnames, MAC and IP addresses
- Git remotes
- Cloud account, database and bucket ids
- Base64 blobs, replaced by their size
Anything the server still recognises as a secret is refused outright: the publish fails rather than storing it. Receipts record which version of the scrubber processed them, so a page can be re-checked when the rules improve.
What still needs your judgement
An optional openai/privacy-filter model runs on your machine and
underlines probable names, private addresses, personal dates and person-identifying
links. It is an aid, not a guarantee: at our calibrated setting it recalled 94%
of planted person names, and it can still miss uncommon names or flag product names,
dates and version strings.
The model never removes text automatically. The preview lists every line
that will be uploaded, marks its guesses with a wavy underline, and lets
you replace each one with [REDACTED]. Employer and client relationships
still require human judgement; only you know whether a name matters.
What you control
- Before publishing: the preview shows the exact outgoing prompt and every receipt line, counts what was removed by category, and lets you edit the prompt, drop individual lines, narrow the time range, or cancel. Publishing is blocked until you confirm you have read them.
- After publishing: the publish token printed by the CLI deletes the page. Keep it; we cannot recover it for you.
- If you spot something later: ask us to redact the receipt, which replaces it with a tombstone and leaves the result standing, or to delete the page entirely.
Rate limiting
To stop abuse we store a keyed hash of the publishing IP address rather than the address. The key changes daily, so the same address does not produce the same hash tomorrow, and rate-limit records older than two days are deleted.
Contact
This is a private preview. It runs on Cloudflare's own hostnames, has no custom domain, and is shared by link rather than listed anywhere. To have something removed or redacted, reply to whoever sent you the link. A monitored takedown address arrives with the domain.