NVIDIA OpenShell

NVIDIA OpenShell is an open-source runtime for autonomous AI agents. It helps developers, platform teams, and security teams control which files, networks, credentials, and model endpoints an agent can use. The runtime enforces those boundaries outside the agent process.

Key Highlights

• Isolated sandboxes with restricted file access and system calls
• Policy-controlled networking with allow and deny logs
• Credentials supplied only to approved endpoints
• A policy prover to check proposed access changes
• Support for Claude Code, Codex, OpenCode, and custom agents

What Makes It Different

OpenShell places enforcement in the execution environment, so access does not depend on an agent following instructions. Its supervisor checks network requests against rules for the calling binary, destination, method, and path.

It is the runtime component of NVIDIA Open Agent Safety Platform. The broader reference design also includes Sentry and BlueField-4; those components are separate from this runtime.

Features & Capabilities

Install the CLI and local gateway, then create a sandbox for your agent. The gateway manages sandbox lifecycle and access. Bring your own agent image: the default minimal Ubuntu image has no agent installed. Docker, Podman, and host virtualization provide local runtime paths, with Helm available for Kubernetes deployments.

Declare access in policy, attach providers, and review requests as the agent works. The policy prover flags risky new access before approval, while the supervisor applies permitted network changes live. Model access can use approved cloud or private endpoints.

User Ratings and Testimonials

For a practical evaluation, test allowed and blocked requests with your own agent and inspect the logs. Windows through WSL 2 is experimental. Docker-backed gateways require Docker 28.0 or later, and Kubernetes needs a network plugin that enforces NetworkPolicy. Sandboxes also require Linux kernel isolation facilities, including Landlock and seccomp, inside containers or VMs.

Pricing & Value

• Open-source runtime: Self-host under the Apache License 2.0, with no runtime license charge.
• Deployment and model access: Supply your own supported infrastructure and configure model providers separately.

Budget for hosting and any model-provider charges alongside the work of maintaining policies. Open-source availability gives teams control over deployment without including hosted infrastructure or model usage.