# AI use cases for Open Shell (NVIDIA/OpenShell) Repo: [NVIDIA/OpenShell](https://github.com/NVIDIA/OpenShell) | Use cases: 3 | Page: https://dedupedai.com/use-cases/nvidia-openshell/ OpenShell is NVIDIA's open-source runtime for running autonomous AI agents safely. Each agent gets an isolated sandbox where kernel controls confine file access and system calls, every network connection passes a policy check, and real credentials are only added to requests bound for approved endpoints. Policy changes are checked with formal verification and risky new access waits for human review. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Run a coding agent in a locked-down sandbox: https://dedupedai.com/use-cases/nvidia-openshell/run-a-coding-agent-in-an-openshell-sandbox/ - Let agents use API keys they can never see: https://dedupedai.com/use-cases/nvidia-openshell/keep-api-keys-away-from-your-agents/ - Teach your coding agent to drive Open Shell: https://dedupedai.com/use-cases/nvidia-openshell/teach-your-agent-to-write-openshell-policies/ ## How to run a coding agent in a locked-down sandbox with Open Shell Repo: [NVIDIA/OpenShell](https://github.com/NVIDIA/OpenShell) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Install Open Shell, create a sandbox and run an agent that must ask before it reaches anything new. **Who it is for:** A developer who wants an agent to work autonomously for hours but not touch the rest of the laptop. **The problem:** Agents are most useful with file, package, API and credential access, which is exactly what makes them risky to leave alone. **The outcome:** An agent working inside an isolated sandbox, with file, system call and network access limited to what a policy allows. ### Steps 1. Run `curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh` to install the CLI and a local gateway 2. Run `openshell sandbox create --name demo` to create your first sandbox 3. Follow the Run Your First Agent guide in the docs, which runs OpenCode against a free OpenRouter model inside the sandbox 4. When the agent needs new access, approve it from Open Shell instead of granting it up front 5. Set `OPENSHELL_TELEMETRY_ENABLED=false` on the gateway if you want to turn off the anonymous telemetry ### Prompt to paste into your agent ```text Install NVIDIA/OpenShell on this machine, create a sandbox named with `openshell sandbox create`, and run inside it on . Every time it asks for new file or network access, show me the request and wait for my approval before allowing it. ``` ### Caveat Open Shell is at 0.1.x and Windows support goes through WSL 2 and is marked experimental, so expect changes between releases. Risk, remote script: The quickstart pipes install.sh from GitHub straight into your shell, so read the script first if you do not want to run it blind. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/nvidia-openshell/run-a-coding-agent-in-an-openshell-sandbox/ ## How to let agents use API keys they can never see with Open Shell Repo: [NVIDIA/OpenShell](https://github.com/NVIDIA/OpenShell) | Difficulty: Advanced | Time: 1 hour | Basis: From the README Give agents credentials that only work at approved endpoints, so a leaked prompt cannot leak the key. **Who it is for:** A team lead whose agents call paid model and SaaS APIs and who worries about keys ending up in logs or the wrong host. **The problem:** An agent that holds a raw API key can send it anywhere, including to a host a prompt injection points it at. **The outcome:** Agents that call approved APIs normally while the real credentials stay outside the sandbox. ### Steps 1. Install Open Shell and create a sandbox as in the quickstart 2. Set up a Provider for each credential; the docs cover providers that work only at approved endpoints, including inference 3. Declare in the sandbox policy which hosts and API methods the agent may reach 4. Let Open Shell add the credential only to requests bound for those approved endpoints; the agent never sees the real value 5. When a policy change would let a credential reach a new host or API method, review what the prover flags before approving it ### Prompt to paste into your agent ```text Using NVIDIA/OpenShell, set up a provider for so the agent in sandbox can call without ever reading the real key. Then try a request to an unapproved host and confirm it is blocked, and show me what the policy prover flags if I add that host. ``` ### Caveat The README points to the docs for provider and policy syntax rather than showing it, so plan to read the Providers and Policies pages. Risk, credentials: You still store the real keys in Open Shell, so protect the gateway like any other secret store and keep each provider scoped to the fewest endpoints. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/nvidia-openshell/keep-api-keys-away-from-your-agents/ ## How to teach your coding agent to drive Open Shell with Open Shell Repo: [NVIDIA/OpenShell](https://github.com/NVIDIA/OpenShell) | Difficulty: Intermediate | Time: 20 minutes | Basis: From the README Install the official Open Shell skills so your agent can run the CLI, write policies and debug gateways. **Who it is for:** A developer who would rather ask Claude Code or Codex to set up sandboxes than learn every policy rule by hand. **The problem:** Sandbox policies and gateway routing are new territory, and a wrong rule either blocks the agent or lets too much through. **The outcome:** A coding agent that can drive the Open Shell CLI, draft sandbox policies and debug gateways and inference routing for you. ### Steps 1. Run `npx skills add NVIDIA/OpenShell` in your project 2. Ask your agent to use the Open Shell skills; they work without an Open Shell source checkout 3. Have it draft a filesystem, network and process policy for your sandbox and review the draft yourself 4. Use the advisor and prover from the Policies docs to check what a change would allow before you apply it 5. For apps, add the Python SDK with `uv add openshell` and use the same release for the SDK and the gateway ### Prompt to paste into your agent ```text Install the NVIDIA/OpenShell skills with `npx skills add NVIDIA/OpenShell`, then write a sandbox policy for that allows and nothing else. Explain each rule, and do not apply it until I approve. ``` ### Caveat An agent-written policy is a draft: the point of Open Shell is that a human reviews new access, so read every rule before you approve it. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/nvidia-openshell/teach-your-agent-to-write-openshell-policies/