OpenShell is NVIDIA's open-source runtime that keeps each AI agent inside a locked-down sandbox, so it can use files, tools and keys without free rein over your data or network.
In plain words
An agent is more helpful when it may open files, add software packages and call APIs, but that is risky. With OpenShell you write a policy saying what each agent may touch. Every agent runs in its own sandbox, and the operating system's kernel checks each file access, system call and network connection against that policy. Agents never see real keys: OpenShell inserts them only into calls headed for services you approved. If a policy change would open new access, a checker flags it for a person to review.
A quick example
Imagine you let a coding agent fix bugs in a private project. Inside OpenShell it can edit the code and download packages from the registry you allowed, but a request to an unknown website is blocked, and your access token is attached only to calls to your approved code host.
Where you'll meet it
- Running coding and research agents with firm limits.
- Letting agents call APIs without showing them secrets.
- Reviewing new access before an agent gets it.
Related
Sources
- DocsNVIDIA/OpenShell, NVIDIA (GitHub) · read 23 Oct 2026
- DocsArchitecture, NVIDIA OpenShell documentation · read 23 Oct 2026
Checked by WhatIsAI against the sources above · 23 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.