Guides
What admins can see
What the admins of your installation can see, what is recorded, what is not, and what even admins cannot do, stated plainly.
App in development
On this page
The portal screens are in development. Their words and layout may still change; the commands behind the buttons are already built.
Admins look after the accounts, data and permissions of your installation. Here is what researchers should know about them, without leaving anything out.
What admins can see
- All source data and all releases. The screens say so: “Admins can always see”.
- Every workspace: the list and states, the data each uses, GPU memory, work-folder use, internet rules and the environment's output.
- Saved work folders and environments. What a release saves is visible to the workspace owner and to admins only. Your Claude conversations (
~/.claude) live in the work folder, so they are part of it. - Approval requests. Admins can see and decide a workspace's requests too.
- The internet access log: which workspace tried to reach where, and whether it was allowed or blocked.
- Account details: name, ID, roles, last sign-in, the list of your SSH keys and API tokens (not their values), and sign-ins that used your correct password without the right code.
- The audit log, described below.
What is recorded
The audit log is append-only by design. It records:
- sign-ins, failed sign-ins and sign-outs;
- commands run from the screen, the API or a workspace, refused and failed ones included;
- creating, opening, starting, stopping and archiving workspaces;
- adding data, approval requests, approvals and rejections;
- releases, and views, previews and downloads of releases;
- file transfers through the IDE;
- permission changes on data, widening who can see a release, and access requests with their outcome;
- connecting Claude, and creating and revoking API tokens and SSH keys;
- GPU memory warnings and processes ended at the limit.
What is not recorded
- Terminals. What you type in a terminal and what it shows are not recorded. The product's own commands run inside a workspace (such as a request to add data) are.
- What goes to the model provider. What the agent sends to the model provider is not recorded by the product. Your Claude conversations simply stay in the work folder.
- SSH. SSH sessions and file transfers are not recorded by the product, only by the server's own OpenSSH logs.
- Older internet access records. The internet access log is a window on recent requests, not part of the audit log.
What even admins cannot do
- Open your IDE. Admins cannot open someone else's workspace IDE; for them, [Open IDE] does not appear on other people's workspaces.
- See token values. No screen shows your Claude token or API tokens again. The Claude token is stored encrypted, and API tokens are not stored in their original form.
- Use admin rights inside a workspace. Inside a workspace there are no admin rights, even when the owner is an admin.
- Connect with your key. On someone else's workspace, the SSH commands say: “These commands connect only with the owner's keys.”
Whoever runs the server
People with root on the server are trusted. They can reach the server's files and records. The audit log is append-only through the database, but root on the server could still change the file itself.
So
- A release puts your Claude conversations into the saved work folder. Keep out of those conversations anything you would not want kept.
- What the agent reads goes to the model provider under your Claude account. Check your Claude privacy settings.
- What the system protects against, and what it does not, is in the security overview.