Nvidia OpenShell

Nvidia's OpenShell introduces a stable release schedule, new isolation features, and expanded APIs in version 0.1.x. This safe, private runtime accommodates fleets of autonomous AI agents while ensuring they have necessary capabilities without unrestricted data access.

Agents require file access, package installation, API calls, and credentials. OpenShell enforces these permissions through policy declarations, preventing unrestricted access to data, secrets, or network. OpenShell governs agent actions via kernel instrumentation for runtime policy enforcement and formal verification of policy changes.

Supported platforms include Linux, macOS on Apple Silicon, and Windows with WSL 2 (experimental), requiring Docker, Podman, or host virtualization. The installer sets up the CLI and local gateway, with a minimal Ubuntu sandbox image included. To run a real agent, follow the Run Your First Agent guide, which demonstrates using OpenCode with a free OpenRouter model and managing new access approvals.

Public OpenShell skills, including coding assistance, policy writing, and gateway debugging, are installable. SDKs facilitate application connections to OpenShell gateways without installing the CLI. Both SDKs and gateways should use the same OpenShell release version when feasible.

OpenShell is agent-first, leveraging the same workflows agents depend on. It employs agent-driven development, development setups, and coding conventions outlined in CONTRIBUTING.md and AGENTS.md. OpenShell gathers minimal anonymous telemetry to aid improvement, avoiding sensitive data like sandbox names, file paths, prompts, credentials, or user content. Users can disable telemetry via environment variables or compile it out entirely.

Retrieved materials, governed by separate terms and conditions, are used at the user's risk. OpenShell is licensed under the Apache License 2.0.