NVIDIA/OpenShell

OpenShell – Secure Runtime for Autonomous AI Agents

What it is – OpenShell is an open‑source platform that lets you run AI agents (e.g., LLM‑powered bots) inside isolated sandboxes. It enforces fine‑grained policies on file access, system calls, and network traffic at the kernel level, and it verifies policy changes formally before they are applied. The goal is to give agents the capabilities they need (install packages, call APIs, use credentials) while keeping your data, secrets, and network safe.

Key components

• Gateway – the control plane that creates sandboxes, stores policies, and mediates access.
• Supervisor – watches each sandbox and applies kernel‑level enforcement.
• Sandbox – an isolated Linux environment (Ubuntu‑based by default) where an agent runs.
• Policy engine – declarative rules for filesystem, network, and process access, with an advisor and prover that flag risky changes.
• Providers – credential bundles that are injected only for approved endpoints (including inference providers).

Supported platforms – Linux, macOS on Apple Silicon, and Windows via WSL 2 (experimental). You can run it locally with Docker/Podman or deploy the gateway on Kubernetes (Helm chart required).

How to try it

# Install the CLI and a local gateway
curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh

# Create a demo sandbox
openshell sandbox create --name demo

The default sandbox is a minimal Ubuntu image with no agent. Follow the “Run Your First Agent” tutorial to launch an OpenCode agent that talks to a free OpenRouter model and demonstrates policy‑approval flows.

SDKs & language bindings – Connect your own applications to an OpenShell gateway with official SDKs:

• Python (uv add openshell)
• TypeScript (npm install @nvidia/openshell-sdk)
• Go (go get github.com/NVIDIA/OpenShell/sdk/go@latest)
• Rust (cargo add openshell-sdk)

Typical use cases

• Deploying LLM‑driven assistants that need to read/write files, install dependencies, or call external APIs.
• Running autonomous coding agents that must fetch secrets only for approved services.
• Providing a safe execution environment for AI‑generated code in enterprise settings.
• Building multi‑tenant AI platforms where each tenant’s agents are isolated by policy.

Extensibility – You can add custom middleware, interceptors, or compute drivers to the sandbox runtime, and you can plug in different inference providers via the Providers abstraction.

Community & support – Discussions on GitHub, issue tracker, security reporting, roadmap project, and a cloud‑hosted launchable on Brev. Telemetry is optional and can be disabled.

License – Apache 2.0.

---

TL;DR – OpenShell gives you a production‑ready, policy‑driven sandbox for running autonomous AI agents safely, with kernel‑level enforcement, formal policy verification, and multi‑language SDKs for integration.