Hacker News
Nvidia OpenShell
SoftwareFirst seen 482 points207 comments
NVIDIA has released OpenShell, an open-source runtime that lets AI agents do useful work while keeping them away from data they should not reach. It is a safe, private space where each agent runs with rules that decide what it can touch.
What happened
NVIDIA published OpenShell on GitHub under the Apache 2.0 open-source license. OpenShell is a runtime, meaning the software environment where an AI agent runs. It lets agents read files, install packages, call APIs, and use credentials. The developer declares what each agent may touch in a policy, and OpenShell enforces that policy. The project also marks version 0.1.x as the start of a stable release schedule.
How it keeps agents safe
OpenShell works in two main ways. First, each agent runs in an isolated sandbox, a sealed box that limits which files and system functions it can use. Every network connection must pass a policy check before it leaves the sandbox. Agents never see real passwords or keys. OpenShell adds them only to requests sent to approved places. Second, before a policy change takes effect, OpenShell uses formal verification, a math-based way to check what new access the change would give. Risky changes, such as giving an agent a new server with credentials, wait for a human to review them.
Key facts
- Works on Linux, macOS on Apple Silicon, and Windows with WSL 2 (experimental).
- Needs Docker, Podman, or host virtualization.
- Installs with one shell command, which sets up a command-line tool and a local gateway.
- The default sandbox is a minimal Ubuntu system with no agent installed.
- Version 0.1.0 includes new isolation tools, a wider extension system, and new programming interfaces (APIs).
Why it matters
AI agents are most helpful when they can act on their own, but that also creates risk. An agent with too much access could leak private data or misuse secrets. OpenShell tries to give agents real power while limiting what they can reach. Its approach of checking policy changes before they apply could help companies trust AI agents with sensitive work.
What to watch
- How well the kernel-level enforcement works in real use.
- Whether the formal verification catches risky policy changes before they cause harm.
- Growth of the project on GitHub and how many developers adopt it.
- Support for more systems, such as full Windows support beyond WSL 2.
Glossary
- OpenShell
- A safe, private runtime from NVIDIA that runs AI agents under rules that limit their access.
- WSL 2
- Windows Subsystem for Linux, a feature that runs a Linux system inside Windows.