Ubuntu and NVIDIA OpenShell: Sandboxed AI Agents from a Snap
Part of our Ubuntu guide
On 1 June 2026, Canonical announced that it would package the NVIDIA OpenShell runtime for agents as a snap on Ubuntu (Canonical announcement). The point of the snap is to make OpenShell a single install command on Ubuntu, with automatic updates and confinement. This article explains what OpenShell does, how the Ubuntu package installs it, and what your machine needs.
What OpenShell does
OpenShell is an open source runtime for autonomous AI agents. It controls what an agent can do with files, networks and tools. Each agent runs in its own isolated sandbox, and you declare what that agent may touch in a policy. The NVIDIA repository describes the enforcement like this: kernel controls limit which files the agent can access and which system calls it can make, and every network connection passes a policy check before it leaves the sandbox.
That is the difference from running an agent directly in a terminal. An agent that runs under your user account can read anything that account can read. In a sandbox, it reaches only what the policy allows. For how the gateway, supervisor and sandbox fit together, the architecture page is the reference.
Why a snap
Canonical's announcement gives three reasons for shipping it as a snap:
- One command to install. The snap bundles the dependencies, so you do not assemble them yourself.
- Automatic updates and channels. The snap receives updates from its channel. The Snap Store lists
latest/stableandlatest/edge. - Strict isolation. The announcement calls it "strict workload isolation" on Ubuntu, which fits a runtime whose job is to isolate other programs.
If you are not sure what a snap is, our comparison of Snap, Flatpak and AppImage explains the three formats.
Install it on Ubuntu
Canonical's instructions are two commands:
sudo snap install openshell # install the NVIDIA OpenShell runtime
openshell sandbox create # start an isolated sandbox for an agent
The Snap Store page lists the package as published by NVIDIA, with a verified publisher account and the Apache-2.0 license. When I checked it, the stable version was 0.1.2 and the listing was updated on 28 September 2026. The version on your machine may differ, so run snap info openshell to see what you have installed.
If snapd is not installed on your system yet, the Ubuntu 26.04 LTS installation guide covers the base system.
What it needs
According to the NVIDIA README, OpenShell runs on Linux, on macOS with Apple Silicon, and on Windows through WSL 2, which the project still labels experimental. It also needs Docker, Podman or host virtualization. The support matrix lists the exact combinations, and the installation page covers the options.
Canonical's announcement names NVIDIA DGX Spark and DGX Station, NVIDIA RTX PRO workstations and NVIDIA DGX data center systems as the target hardware. The announcement does not say which hardware is required beyond that, so check the support matrix before you plan a deployment.
Where it fits
OpenShell runs the agent. It does not replace the agent. Tools that use it already exist: the NVIDIA NemoClaw repository, for example, describes running agents such as Hermes, LangChain Deep Agents and OpenClaw inside OpenShell.
If you already run agents from a terminal, a multiplexer keeps them organised and a sandbox limits what they can reach. They solve different problems. Our guide to running several Claude Code agents in one terminal covers the first. For agents on a server with no desktop at all, running AI agents without a GUI covers the systemd side.
What this article does not cover
The Canonical announcement describes the snap and the install commands. It does not describe how to write a policy, which agents work with it, or how it performs. Those details are in the NVIDIA documentation, and this article does not repeat them. Before you put an agent behind a policy you wrote yourself, read the NVIDIA docs and test the policy on a throwaway sandbox.
JM Carreira
I’ve used Linux since 2008 and it has been my main system since 2021. When an article has a “Tested on” box, I ran the steps on my Lenovo Tiny. More about me and the test machine.