How we protect your keys, papers and work

• Encrypted keys

  In place

  Your AI provider keys and two-factor secrets are encrypted before they are stored, each kind under its own separate key. Once saved, a provider key is only ever shown as its last four characters. It is never sent back to your browser or written to logs.

• Uploaded papers

  In place

  Text from files you upload is only ever read by Anthropic's models. Other AI providers see a placeholder instead. Once an agent has read one of your files, it can no longer search the web, so nothing from the file can leak out in a search. Every file is checked before it is stored: PDFs that are encrypted or contain scripts are rejected, and the rest are saved as a clean copy. Two limits to know about: what an agent writes in its own words after reading your file isn't tracked, and files aren't virus-scanned.

• Activity logs

  In place

  Detailed activity logging is off by default. If your administrator turns it on, your prompts and the agents' answers are sent to your administrator's logging service (Langfuse), with API keys and file contents hidden. Other personal details are not hidden.

• Internal network

  In place

  Only the front door of the service is reachable from the internet. The database and file storage sit on an internal network that the outside world can't reach.

• Private file links

  In place

  Uploads and exports are kept in private storage, separate for each workspace. Your browser never talks to that storage directly: our server fetches the file using a link that expires within minutes and passes it on to you as a download.

• Two-factor authentication

  In place

  Sign in with a code from an authenticator app as well as your password. After too many wrong codes, sign-in is paused for a while.

• Audit log

  In place

  Important actions are recorded with who did them and when: sign-ins and sign-ups, two-factor setup, key changes, workspace limit changes, uploads and deletions, exports and downloads, batches, spending decisions, and every review, approval and rejection.

• Spending limits

  In place

  You use your own AI provider key, so you stay in control of the bill. Each session has a warning limit and a hard limit, set for your workspace. Work already in progress may finish slightly past a limit; then you choose whether to raise it or wrap up. Plans and batches show a cost estimate before anything runs, and a batch stops starting new targets if spending runs well past the estimate. You can stop everything at any time.

• Untrusted content

  In place

  Web pages, uploaded files and tool results are clearly marked as outside content before an agent reads them, so the agents treat them as information, not instructions. Everything shown in the app and every export is cleaned first. This lowers the risk of a document trying to trick the agents; it can't remove that risk entirely.

• Restricted internet access

  In place

  The research tools can only reach an approved list of websites and databases. Web search results also follow your workspace's own lists of blocked and allowed sites.

• Workspace isolation

  In place

  Every request is checked on our server to confirm which workspace it belongs to. Projects, targets, conversations, files, exports and keys all belong to one workspace, and no workspace can see another's data.

• Password reset and account recovery

  In progress

  Password reset, two-factor recovery and limits on repeated sign-in attempts are still being built. Until then, if you lose your authenticator app, an administrator has to reset it for you.

Join the waitlistRead our privacy policy