An AI agent only saves you work if it can read files, install things and use your passwords and access. That same access lets it leak your data or send it where it shouldn't go. OpenShell locks each agent in an isolated environment and gives it a written list of the files, sites and credentials it may use; everything else is blocked.
Who this is NOT for
If you use AI in the browser (ChatGPT, Claude on the web) and no agent touches your business files or accounts, there is nothing to protect here: do not install it. The same goes if nobody on your team works in the terminal or writes YAML rules: the isolated environment comes empty, with no agent, and setting it up is technical work. And if your business runs on Windows, note that it only works there through WSL 2, as an experimental feature.
The obvious move is to drop the agent in a Docker container and call it done. The difference is that Docker isolates it but does not decide which sites it may reach or with which credentials. Here the agent never sees your real keys: OpenShell adds them only when the request goes to an approved destination. And before you approve a new permission, it warns you what risky access it would open. Something we learned running agents: an agent that is only 'restricted' in its own configuration is not actually restricted, the limit has to live outside it. Watch two things. It is a young project (created February 2026, version 0.1.x, 576 open issues and PRs), though NVIDIA updates it daily. And it sends anonymous telemetry by default; if you don't want that, turn it off with OPENSHELL_TELEMETRY_ENABLED=false. The typical mistake is approving new permissions without reading them: approve everything out of fatigue and the cage is open.
Your first step
On a test machine (not your production server), run curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh and then openshell sandbox create --name demo. Then follow the official "Run Your First Agent" guide to see how you approve each new access.
The OpenShell implementation guide
The project belongs to NVIDIA Corporation (NVIDIA) and it is free: the link beside asks for nothing. The guide to set it up without getting lost is ours: step by step, common mistakes and a checklist. Leave your email and we will send it.
No spam. Unsubscribe anytime.