NVIDIA OpenShell

NVIDIA OpenShell is an open-source runtime for running AI agents inside environments with explicit access policies. It is useful when a coding agent needs a working directory and model access, but should not inherit everything available to your login session. NVIDIA announced the project on March 16, 2026; this entry was checked against the current documentation on September 28.

Key features

• Filesystem boundaries: define which paths an agent can read and change.
• Network policies: restrict outbound destinations and review requests for additional access.
• Provider profiles: associate credentials with permitted endpoints and executables.
• Reusable environments: run an agent from an OCI image containing its required tools.
• Policy review: keep YAML policies alongside project configuration so changes can be inspected.

The runtime supports coding tools such as Claude Code, OpenCode, Codex, and GitHub Copilot CLI. It supplies the execution boundary, while the selected agent still performs the work. See the official overview.

A practical workflow

Start with a small repository copy and a task such as generating documentation. Identify the source files the agent needs, the model endpoint it will use, and whether package downloads are necessary. Configure those permissions before running the task.

When something is blocked, inspect the denied destination and calling program. Approve a narrow rule only if it is needed for the task. A useful acceptance check is to verify both sides: permitted project files should work, while an unrelated directory and an unapproved endpoint should remain inaccessible. Finally, review the resulting changes and preserve the policy with the project.

Quick start and requirements

1. Check the support matrix for your host and chosen runtime.
2. Install a stable CLI release using the official installation guide.
3. Configure a provider profile and choose an image containing your agent.
4. Follow Run Your First Agent, then test a limited task.

Linux and Apple Silicon macOS are supported in the documented configurations; Windows through WSL 2 is experimental. Kernel and runtime requirements matter, so a successful container installation alone does not establish compatibility.

Pricing and limitations

OpenShell is Apache-2.0 software. Model API usage, compute, storage, and operation costs remain separate. An allowed action can still be harmful or incorrect: permitting writes to a repository does not guarantee a correct patch. Use backups and normal code review. This listing is based on documentation, not an independent security audit.

FAQ

Does OpenShell replace a coding agent?

No. You provide the agent and its tools in an image; OpenShell governs their execution environment.

Does a sandbox make cloud inference local?

No. Your selected provider and routing determine where inference runs. Review the provider configuration before sending sensitive context.

Alternatives

• E2B: compare when your application needs sandbox execution through an SDK.
• Daytona: compare for programmatic development environments.

Explore the Agent Infra category and sandbox tools. OpenShell is worth evaluating when explicit file, network, and credential boundaries are part of your agent deployment requirements.