How to try an AI agent in a test environment
Before you let an AI agent work with your folders, you can try it apart and see what it could touch. This guide doesn't promise a foolproof environment: it helps you set up one where you can see that.
Why an empty folder isn't enough
Opening the agent in an empty folder doesn't isolate it. Most agents can read your whole home folder, your keys too, and many can go online.
Three ways to isolate it
- With its own profile. Some agents have a setting that stops them reading outside the project. In Claude Code it's the Read only profile, which RosmarOps's report gives you.
- With its sandbox. Many agents have one, almost always off and often only on macOS and Linux. The table shows which one yours has and whether it's on. Note: many limit what it writes and where it connects, but not what it reads.
- With another user account. If your agent has no other way, create a new user account on your computer, without your folders or keys, and open it there.
Then: see what it could touch
Run RosmarOps in that same environment. It tells you what the agent can do without asking you, with its control score.
What this guide doesn't do
RosmarOps doesn't set up the environment or change your agent's settings: it only reads. A test environment limits what the agent could touch; it doesn't make it harmless.
