Secure Long Running AI Agents with OpenShell on DGX Station

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.

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

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.

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.

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

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}'

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

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

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).

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).

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

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.

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

Configure OpenClaw within OpenShell Sandbox

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

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.

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!"
      }'

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.

Reconnect to the sandbox

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

openshell sandbox connect dgx-demo

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

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

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

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.