dgx-spark-playbooks/nvidia/station-openshell/README.md

Secure Long Running AI Agents with OpenShell on DGX Station

Run OpenClaw with local models in an NVIDIA OpenShell sandbox on DGX Station

Table of Contents

• Overview
  • Notice & Disclaimers
• Instructions
  • 5a. Start the OpenAI-compatible server on port 8000
  • Access the dashboard
• Troubleshooting

---

Overview

Basic idea

OpenClaw is a local-first AI agent that runs on your machine, combining memory, file access, tool use, and community skills into a persistent assistant. Running it directly on your system means the agent can access your files, credentials, and network—creating real security risks.

NVIDIA OpenShell solves this problem. It is an open-source sandbox runtime that wraps the agent in kernel-level isolation with declarative YAML policies. OpenShell controls what the agent can read on disk, which network endpoints it can reach, and what privileges it has—without disabling the capabilities that make the agent useful.

By combining OpenClaw with OpenShell on DGX Station (with NVIDIA GB300 GPUs), you get the full power of a local AI agent backed by GPU memory for local models, while enforcing explicit controls over filesystem access, network egress, and credential handling.

Notice & Disclaimers

Quick Start Safety Check

Use a clean environment only. Run this playbook on a fresh device or VM with no personal data, confidential information, or sensitive credentials. Think of it like a sandbox—keep it isolated.

By installing this playbook, you're taking responsibility for all third-party components, including reviewing their licenses, terms, and security posture. Read and accept before you install or use.

---

What You're Getting

The playbook showcases experimental AI agent capabilities. Even with cutting-edge open-source tools like OpenShell in your toolkit, you need to layer in proper security measures for your specific threat model.

---

Key Risks with AI Agents

Be mindful of these risks with AI agents:

1. Data leakage – Any materials the agent accesses could be exposed, leaked, or stolen.
2. Malicious code execution – The agent or its connected tools could expose your system to malicious code or cyber-attacks.
3. Unintended actions – The agent might modify or delete files, send messages, or access services without explicit approval.
4. Prompt injection & manipulation – External inputs or connected content could hijack the agent's behavior in unexpected ways.

---

Security Best Practices

No system is perfect, but these practices help keep your information and systems safe:

1. Isolate your environment – Run on a clean PC or isolated virtual machine. Only provision the specific data you want the agent to access.
2. Never use real accounts – Don't connect personal, confidential, or production accounts. Create dedicated test accounts with minimal permissions.
3. Vet your skills/plugins – Only enable skills from trusted sources that have been vetted by the community.
4. Lock down access – Ensure your OpenClaw UI or messaging channels aren't accessible over the network without proper authentication.
5. Restrict network access – Where feasible, limit the agent's internet connectivity.
6. Clean up after yourself – When you're done, remove OpenClaw and revoke all credentials, API keys, and account access you granted.

---

What you'll accomplish

You will install the OpenShell CLI (openshell), deploy a gateway on your DGX Station, and launch OpenClaw inside a sandboxed environment using the pre-built OpenClaw community sandbox. The sandbox enforces filesystem, network, and process isolation by default. You will also configure local inference routing so OpenClaw uses a model running on your DGX Station via vLLM (NVIDIA NGC container on the host) without needing external API keys.

Popular use cases

• Secure agent experimentation: Test OpenClaw skills and integrations without exposing your main filesystem or credentials to the agent.
• Private enterprise development: Route all inference to a local model on DGX Station. No data leaves the machine unless you explicitly allow it in the policy.
• Auditable agent access: Version-control the policy YAML alongside your project. Review exactly what the agent can reach before granting access.
• Iterative policy tuning: Monitor denied connections in real time with openshell term, then hot-reload updated policies without recreating the sandbox.

What to know before starting

• Comfort with the Linux terminal and SSH
• Basic understanding of Docker (OpenShell runs a k3s cluster inside Docker)
• Familiarity with Docker and local LLM serving (vLLM in this playbook)
• Awareness of the security model: OpenShell reduces risk through isolation but cannot eliminate all risk. Review the OpenShell overview and OpenClaw security guidance.

Prerequisites

Hardware Requirements:

• NVIDIA DGX Station with GB300 GPU(s)
• Sufficient GPU memory for your chosen model: we recommend Nemotron 3 Super in NVFP4 (nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4) served by vLLM on GB300; smaller GPUs can use other Hugging Face models—check nvidia-smi and model requirements

Software Requirements:

• DGX OS or Ubuntu 24.04 (or compatible Linux)
• Docker Desktop or Docker Engine running: docker info
• Python 3.12 or later: python3 --version
• uv package manager: uv --version (install with curl -LsSf https://astral.sh/uv/install.sh | sh)
• NVIDIA Container Toolkit (for GPU-enabled Docker): configured per instructions.md
• Network access to download Python packages from PyPI, the NGC vLLM image, and model weights from Hugging Face

Time & risk

• Estimated time: 20–30 minutes (plus model download time, which depends on model size and network speed).
• Risk level: Low to Medium
  • OpenShell sandboxes enforce kernel-level isolation, significantly reducing the risk compared to running OpenClaw directly on the host.
  • The sandbox default policy denies all outbound traffic not explicitly allowed. Misconfigured policies may block legitimate agent traffic; use openshell logs to diagnose.
  • Large model downloads may fail on unstable networks.
• Rollback: Delete the sandbox with openshell sandbox delete dgx-demo (or your sandbox name), stop the gateway with openshell gateway stop, and optionally destroy it with openshell gateway destroy. Stop and remove the vLLM container and delete Hugging Face cache directories if you want to reclaim disk space (see instructions.md cleanup).
• Last Updated: 03/13/2026

Instructions

Step 1. Confirm your environment

Verify the OS, GPU, Docker, and Python are available before installing anything.

head -n 2 /etc/os-release
nvidia-smi
docker info --format '{{.ServerVersion}}'
python3 --version

Expected output should show Ubuntu 24.04 (DGX OS), a detected GPU (e.g. NVIDIA GB300 on DGX Station), a Docker server version, and Python 3.12+. If you access the DGX Station remotely, ensure port 18789 is available for the OpenClaw dashboard.

Step 2. Docker Configuration

First, verify that the local user has Docker permissions using the following command.

docker ps

If you get a permission denied error (permission denied while trying to connect to the docker API at unix:///var/run/docker.sock), add your user to the system's Docker group. This will enable you to run Docker commands without requiring sudo. The command to do so is as follows:

sudo usermod -aG docker $USER
newgrp docker

Now that we have verified the user's Docker permission, we must configure Docker so that it can use the NVIDIA Container Runtime.

sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

Run a sample workload to verify the setup:

docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi

Step 3. Install the OpenShell CLI

Create a virtual environment and install the openshell CLI.

cd ~
uv venv openshell-env && source openshell-env/bin/activate
uv pip install openshell 
openshell --help

If you don't have uv installed yet:

curl -LsSf https://astral.sh/uv/install.sh | sh
export PATH="$HOME/.local/bin:$PATH"

Expected output should show the openshell command tree with subcommands like gateway, sandbox, provider, and inference.

Step 4. Deploy the OpenShell gateway on DGX Station

The gateway is the control plane that manages sandboxes. Since you are running directly on the DGX Station, it deploys locally inside Docker.

openshell gateway start
openshell status

openshell status should report the gateway as Connected. The first run may take a few minutes while Docker pulls the required images and the internal k3s cluster bootstraps.

Note

Remote gateway deployment requires passwordless SSH access. Ensure your SSH public key is added to ~/.ssh/authorized_keys on the DGX Station before using the --remote flag.

Tip

If you want to manage the DGX Station gateway from a separate workstation, run openshell gateway start --remote <username>@<dgx-station-ip-or-hostname> from that workstation instead. All subsequent commands will route through the SSH tunnel.

Step 5. Run vLLM with Nemotron 3 Super (recommended)

vLLM path only.

docker pull nvcr.io/nvidia/vllm:26.03-py3

5a. Start the OpenAI-compatible server on port 8000

The OpenShell gateway must reach this service using the host’s real IP address (not localhost from inside other containers). Binding --host 0.0.0.0 and publishing -p 8000:8000 makes the API available on all interfaces.

The Nemotron weights may require a Hugging Face account and token. Create your own read token at huggingface.co/settings/tokens, keep it private (do not paste real tokens into shared docs, tickets, or git), then export it in your shell before docker run so the command below only references the variable:

export HF_TOKEN=your_actual_token_here

Replace your_actual_token_here with your real token value. If you do not need Hugging Face authentication for this model, skip the export and remove the -e HF_TOKEN="$HF_TOKEN" line from the docker run command.

We are going to use the nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 model as it fits in DGX Station VRAM with KV headroom at --max-model-len 32768

Warning

The --trust-remote-code flag in the following docker run command allows execution of arbitrary code from the model repository. Only use this with trusted models.

docker run -d --name vllm-openshell \
  --runtime nvidia --gpus all \
  -e HF_TOKEN="$HF_TOKEN" \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 \
  --restart unless-stopped \
  nvcr.io/nvidia/vllm:26.03-py3 \
  python3 -m vllm.entrypoints.openai.api_server \
    --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 \
    --host 0.0.0.0 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --trust-remote-code \
    --max-model-len 32768 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_xml \
    --reasoning-parser nemotron_v3

Watch logs until the server is ready (first start can take several minutes while weights load). Then, in a new terminal window, run:

docker logs -f vllm-openshell

Wait for logs to output Application startup complete., then verify the API using:

curl -s http://localhost:8000/v1/models

You should see JSON listing nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4.

Warm up with a short completion so CUDA graphs compile before OpenClaw validates the route (first request may take 30–90 seconds):

curl -s --max-time 120 http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4","messages":[{"role":"user","content":"Say hello."}],"max_tokens":16}'

Step 6. Create an inference provider

Create an OpenShell provider that points at the vLLM OpenAI-compatible API on the host (/v1 on port 8000).

First, find the IP address of your DGX Station:

hostname -I | awk '{print $1}'

Then create the provider, replacing {Machine_IP} with the IP address from the command above (e.g. 10.110.106.169):

openshell provider create \
    --name local-vllm \
    --type openai \
    --credential OPENAI_API_KEY=not-needed \
    --config OPENAI_BASE_URL=http://{Machine_IP}:8000/v1

Important

Do not use localhost or 127.0.0.1 here. The OpenShell gateway runs inside a Docker container, so it cannot reach the host via localhost. Use the machine's actual IP address.

Some Linux Docker setups can use http://host.docker.internal:8000/v1 instead of the host IP; if your gateway resolves that hostname, it is equivalent.

Verify the provider was created:

openshell provider list

Step 7. Configure inference routing

Point the inference.local endpoint (available inside every sandbox) at vLLM. The model id must match what /v1/models returns (for the default Step 5 command, use the Hugging Face id below). If you changed --model in Step 5, use that same string here.

openshell inference set \
    --provider local-vllm \
    --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4

The output should confirm the route and show a validated endpoint URL, for example: http://10.110.106.169:8000/v1/chat/completions (openai_chat_completions).

Note

If you see failed to verify inference endpoint or failed to connect, ensure vLLM is healthy (docker logs vllm-openshell) and you completed at least one chat completion so cold-start compilation has finished. You can add --no-verify to skip verification: openshell inference set --provider local-vllm --model nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 --no-verify.

Verify the configuration:

openshell inference get

Expected output should show provider: local-vllm and model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 (or whichever model you configured in Step 5).

Step 8. Deploy OpenShell Sandbox

Create a sandbox using the pre-built OpenClaw community sandbox. This pulls the OpenClaw Dockerfile, the default policy, and startup scripts from the OpenShell Community catalog:

openshell sandbox create \
  --keep \
  --forward 18789 \
  --name dgx-demo \
  --from openclaw \
  -- openclaw-start

Note

Do not pass --policy with a local file path (e.g. openclaw-policy.yaml) when using --from openclaw. The policy is bundled with the community sandbox; a local file path can cause "file not found."

The --keep flag keeps the sandbox running after the initial process exits, so you can reconnect later. This is the default behavior. To terminate the sandbox when the initial process exits, use the --no-keep flag instead.

Note

The sandbox name is displayed in the creation output. You can also set it explicitly with --name <your-name>. To find it later, run openshell sandbox list.

The CLI will:

1. Resolve openclaw against the community catalog
2. Pull and build the container image
3. Apply the bundled sandbox policy
4. Launch OpenClaw inside the sandbox

Step 9. Configure OpenClaw within OpenShell Sandbox

The sandbox container will spin up and the OpenClaw onboarding wizard will launch automatically in your terminal.

Important

The onboarding wizard is fully interactive — it requires arrow-key navigation and Enter to select options. It cannot be completed from a non-interactive session (e.g. a script or automation tool). You must run openshell sandbox create from a terminal with full TTY support.

If the wizard did not complete during sandbox creation, reconnect to the sandbox to re-run it:

openshell sandbox connect dgx-demo

Use the arrow keys and Enter key to interact with the installation.

• If you understand and agree, use the arrow key of your keyboard to select 'Yes' and press the Enter key.
• Quickstart vs Manual: select Quickstart and press the Enter key.
• Model/auth Provider: Select Custom Provider, the second-to-last option.
• API Base URL: update to https://inference.local/v1
• How do you want to provide this API key?: Paste API key for now.
• API key: enter not-needed (or any placeholder; vLLM is not checking the key unless you enabled API-key auth in the server).
• Endpoint compatibility: select OpenAI-compatible and press Enter.
• Model ID: enter the same id you set in Step 7 (e.g. nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4).
  • The first request may take up to a minute while vLLM compiles; ensure the container from Step 5 is already serving (docker logs vllm-openshell).
• Endpoint ID: leave the default value.
• Alias: enter the same model name (this is optional).
• Channel: Select Skip for now.
• Search provider: Select Skip for now.
• Skills: Select No for now.
• Enable hooks: Select No for now (using the space bar) and press Enter.

It might take 1-2 minutes to get through the final stages. Afterwards, you should see a URL with a token you can use to connect to the gateway.

The expected output will be similar, but the token will be unique.

OpenClaw gateway starting in background.
  Logs: /tmp/gateway.log
  UI:   http://127.0.0.1:18789/?token=9b4c9a9c9f6905131327ce55b6d044bd53e0ec423dd6189e

In order to verify the default policy enabled for your sandbox, please run the following command:

openshell sandbox get dgx-demo

Access the dashboard

Accessing the dashboard from DGX station as the primary device: right-click on the URL in the UI section and select Open Link.

Accessing the dashboard from the host or a remote system: The dashboard URL (e.g. http://127.0.0.1:18789/?token=...) is inside the sandbox, so the host does not forward port 18789 by default. To reach it from your host or another machine, use SSH local port forwarding. From a machine that can reach the OpenShell gateway, run (replace gateway URL, sandbox-id, token, and gateway-name with values from your environment):

ssh -o ProxyCommand='/usr/local/bin/openshell ssh-proxy --gateway https://127.0.0.1:8080/connect/ssh --sandbox-id <sandbox-id> --token <token> --gateway-name openshell' -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null -o LogLevel=ERROR -N -L 18789:127.0.0.1:18789 sandbox

Then open http://127.0.0.1:18789/?token=<your-token> in your local browser.

To access from another machine, use the SSH tunnel described above, or open the dashboard URL in your browser (e.g. after port forwarding or from the DGX Station's local browser).

From this page, you can now Chat with your OpenClaw agent within the protected confines of the runtime OpenShell provides.

Step 10. Conduct Inference within Sandbox

Connecting to the Sandbox (Terminal)

Now that OpenClaw has been configured within the OpenShell protected runtime, you can connect directly into the sandbox environment via:

openshell sandbox connect dgx-demo

Once loaded into the sandbox terminal, you can test connectivity to vLLM via inference.local with this command:

curl https://inference.local/v1/responses \
          -H "Content-Type: application/json" \
          -d '{
        "instructions": "You are a helpful assistant.",
        "input": "Hello!"
      }'

Step 11. Verify sandbox isolation

Open a second terminal and check the sandbox status and live logs:

source ~/openshell-env/bin/activate
openshell term

The terminal dashboard shows:

• Sandbox status — name, phase, image, providers, and port forwards
• Live log stream — outbound connections, policy decisions (allow, deny, inspect_for_inference), and inference interceptions

Verify that the OpenClaw agent can reach inference.local for model requests and that unauthorized outbound traffic is denied.

Tip

Press f to follow live output, s to filter by source, and q to quit the terminal dashboard.

Step 12. Reconnect to the sandbox

If you exit the sandbox session, reconnect at any time:

openshell sandbox connect dgx-demo

Note

openshell sandbox connect is interactive-only — it opens a terminal session inside the sandbox. There is no way to pass a command for non-interactive execution. Use openshell sandbox upload/download for file transfers, or use the SSH proxy for scripted access (see Step 9).

To transfer files in or out (replace dgx-demo with your sandbox name if you used a different one):

openshell sandbox upload dgx-demo ./local-file /sandbox/destination
openshell sandbox download dgx-demo /sandbox/file ./local-destination

Step 13. Cleanup

Stop and remove the sandbox (use the name you gave it, e.g. dgx-demo):

openshell sandbox delete dgx-demo

Stop the gateway (preserves state for later):

openshell gateway stop

Warning

The following command permanently removes the gateway cluster and all its data.

openshell gateway destroy

Remove the inference provider you created in Step 6:

openshell provider delete local-vllm

Stop and remove the vLLM container started in Step 5:

docker stop vllm-openshell
docker rm vllm-openshell

(Optional) Remove the container image to free disk:

docker rmi nvcr.io/nvidia/vllm:26.03-py3

Step 14. Next steps

• Add more providers: Attach GitHub tokens, GitLab tokens, or cloud API keys as providers with openshell provider create. When creating the sandbox, pass the provider name(s) with --provider <name> (e.g. --provider my-github) to inject those credentials into the sandbox securely.
• Try other community sandboxes: Run openshell sandbox create --from base or --from sdg for other pre-built environments.
• Connect VS Code: Use openshell sandbox ssh-config <sandbox-name> and append the output to ~/.ssh/config to connect VS Code Remote-SSH directly into the sandbox.
• Monitor and audit: Use openshell logs <sandbox-name> --tail or openshell term to continuously monitor agent activity and policy decisions.

Troubleshooting

Symptom	Cause	Fix
openshell gateway start fails with "connection refused" or Docker errors	Docker is not running	Start Docker with sudo systemctl start docker or launch Docker Desktop, then retry openshell gateway start
openshell status shows gateway as unhealthy	Gateway container crashed or failed to initialize	Run openshell gateway destroy and then openshell gateway start to recreate it. Check Docker logs with docker ps -a and docker logs <container-id> for details
openshell sandbox create --from openclaw fails to build	Network issue pulling the community sandbox or Dockerfile build failure	Check internet connectivity. Retry the command. If the build fails on a specific package, check if the base image is compatible with your Docker version
Sandbox is in Error phase after creation	Policy validation failed or container startup crashed	Run openshell logs <sandbox-name> to see error details. Common causes: invalid policy YAML, missing provider credentials, or port conflicts
Agent cannot reach inference.local inside the sandbox	Inference routing not configured or provider unreachable	Run openshell inference get to verify the provider and model are set. From the host, test vLLM: curl -s http://localhost:8000/v1/models. The provider base URL must use the host’s real IP (not 127.0.0.1/localhost) so the gateway container can reach vLLM (see instructions.md Step 6).
503 verification failed or timeout when the gateway validates vLLM	vLLM not listening on all interfaces, firewall blocking port 8000, model still loading, or first-request CUDA graph compile	Ensure the vLLM server was started with --host 0.0.0.0 and port 8000 mapped (see Step 5). Warm up with a chat completion request before openshell inference set. Allow port 8000 if you use a host firewall: sudo ufw allow 8000/tcp comment 'vLLM for OpenShell Gateway' (then sudo ufw reload if needed). For very large models, try openshell inference set ... --no-verify after confirming vLLM works from the host.
Agent's outbound connections are all denied	Default policy does not include the required endpoints	Monitor denials with openshell logs <sandbox-name> --tail --source sandbox. Pull the current policy with openshell policy get <sandbox-name> --full, add the needed host/port under network_policies, and push with openshell policy set <sandbox-name> --policy <file> --wait
"Permission denied" or Landlock errors inside the sandbox	Agent trying to access a path not in read_only or read_write filesystem policy	Pull the current policy and add the path to read_write (or read_only if read access is sufficient). Push the updated policy. Note: filesystem policy is static and requires sandbox recreation
vLLM OOM or very slow inference	Model too large for available VRAM, --max-model-len too high, or GPU contention	Free GPU memory (close other GPU workloads), use a smaller Hugging Face model or quantized variant, or lower --max-model-len. Check docker logs for the vLLM container. Monitor with nvidia-smi
openshell sandbox connect hangs or times out	Sandbox not in Ready phase	Run openshell sandbox get <sandbox-name> to check the phase. If stuck in Provisioning, wait or check logs. If in Error, delete and recreate the sandbox
Policy push returns exit code 1 (validation failed)	Malformed YAML or invalid policy fields	Check the YAML syntax. Common issues: paths not starting with /, .. traversal in paths, root as run_as_user, or endpoints missing required host/port fields. Fix and re-push
openshell gateway start fails with "K8s namespace not ready" / timed out waiting for namespace	The k3s cluster inside the Docker container takes longer to bootstrap than the CLI timeout allows. The internal components (TLS secrets, Helm chart, namespace creation) may need extra time, especially on first run when images are pulled inside the container.	First, check whether the container is still running and progressing: docker ps --filter name=openshell (look for health: starting). Inspect k3s state inside the container: docker exec <container> sh -c "KUBECONFIG=/etc/rancher/k3s/k3s.yaml kubectl get ns" and kubectl get pods -A. If pods are in ContainerCreating and TLS secrets are missing (navigator-server-tls, openshell-server-tls), the cluster is still bootstrapping — wait a few minutes and run openshell status again. If it does not recover, destroy with openshell gateway destroy (and docker rm -f <container> if needed) and retry openshell gateway start. Ensure Docker has enough resources (memory and disk) for the k3s cluster.
openshell status says "No gateway configured" even though the Docker container is running	The gateway start command failed or timed out before it could save the gateway configuration to the local config store	The container may still be healthy — check with docker ps --filter name=openshell. If the container is running and healthy, try openshell gateway start again (it should detect the existing container). If the container is unhealthy or stuck, remove it with docker rm -f <container> and then openshell gateway destroy followed by openshell gateway start.