Security

Your agents get real access to your code and logins. Cloudroom is built to keep that access contained.

How we protect your work

Each cloud thread runs in its own isolated sandbox. Sandboxes are never shared between threads or users.
Agents get full admin rights inside their own sandbox, but no access to other sandboxes.
The access token Cloudroom keeps in a sandbox only works for that one sandbox.
Every request to your core needs a secret 256-bit token. The desktop app only connects to remote cores over HTTPS.
Session history is saved outside the sandbox, over an encrypted connection with full certificate checks.
Your agents call your model providers directly, with your own logins. Cloudroom never sits in between.
When an agent needs an API key, you enter it in a secure form. The key never appears in the chat, history, or logs.
Command Guard, on by default, stops a few catastrophic commands before they run, such as deleting your home folder, formatting a disk, or deleting a GitHub repository.
The Cloudroom core is open source under Apache 2.0, so anyone can audit it.

Agents run without approval prompts. With admin rights, they can read anything in their sandbox, including the keys Cloudroom keeps there.
Subagents share their thread’s sandbox, so they can change each other’s files.
With Mac access on, cloud agents can read files and run commands on your Mac as you. You can turn it off anytime in Settings.
Command Guard only catches a few known commands, and you can turn it off in Settings.

Email david@davidondrej.com or report it privately on GitHub. Please do not open public issues for security problems.