Agent Sandbox

OpenClaw Sandbox Example

This example demonstrates how to run OpenClaw (formerly Moltbot) inside the Agent Sandbox.

Prerequisites

• A Kubernetes cluster (e.g., Kind).
• (Optional) Ensure your cluster has a RuntimeClass (e.g., gvisor) configured and nodes support it. This example is verified using gVisor. The openclaw-sandbox.yaml manifest includes a commented-out runtimeClassName: gvisor line. Uncomment it or update it if you are using a non-default runtime class (e.g., Kata Containers). See the gVisor documentation.
• agent-sandbox controller installed.

Usage

1. (If using Kind) Load the image into Kind:

   kind load docker-image ghcr.io/openclaw/openclaw:2026.3.23
   

2. Generate a secure token:

   export OPENCLAW_GATEWAY_TOKEN="$(openssl rand -hex 32)"
   

3. Apply the Sandbox resource (replacing the token placeholder):

   kubectl apply -f openclaw-config.yaml
   
   sed "s/dummy-token-for-sandbox/$OPENCLAW_GATEWAY_TOKEN/g" openclaw-sandbox.yaml | kubectl apply -f -
   

4. Access the Web UI:

   Option 1: Direct Port-Forward (Only if NOT using gVisor) Verify the pod is running and port-forward to access it directly:

   kubectl port-forward pod/openclaw-sandbox 18789:18789
   

   Then open http://localhost:18789 in your browser.

   Option 2: Access with gVisor Enabled If you enable gVisor by uncommenting runtimeClassName: gvisor in openclaw-sandbox.yaml, direct kubectl port-forward to the pod will fail (see Issue #158).

   To access the Web UI with gVisor, you must use an alternative method:

   • Kubernetes Service: Expose the sandbox pod via a NodePort or LoadBalancer service and access it via the service’s endpoint.
   • Router Architecture: Use the sandbox-router to proxy traffic. See agentic-sandbox-client and sandbox-router for instructions.

CLI Operations

You can run OpenClaw CLI commands directly inside the sandbox container.

kubectl exec -it openclaw-sandbox -- openclaw --help