Secure AI Agents with OpenShell
Isolate OpenClaw with kernel-level policies and route inference to a local model
Agent-ready models
Agent-ready models are tuned for agentic workloads — tool calling, reasoning traces, and long multi-turn sessions. Pick the recommended model for your hardware platform, serve it with an OpenAI-compatible API (this playbook uses vLLM), then continue with the Instructions tab to wire OpenShell inference routing and launch the OpenClaw sandbox.
Recommendations by hardware platform
| Hardware platform | Recommended agent-ready model | HuggingFace handle | Notes |
|---|---|---|---|
| DGX Spark | Agent-ready Qwen3.6-35B-A3B (NVFP4) | nvidia/Qwen3.6-35B-A3B-NVFP4 | Recommended for unified-memory agentic serving; launch via Agent-ready Models |
| DGX Station | NVIDIA Nemotron 3 Ultra | nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 | Recommended for Station; launch via the DGX Station deployment guide |
NOTE
DGX Spark: launch via Agent-ready Models. DGX Station: use the Nemotron 3 Ultra Station deployment guide. OpenShell requires the server to listen on all interfaces (--host 0.0.0.0) so the gateway container can reach it — keep that flag when you launch.
Before you serve
- Complete Docker setup in the Instructions tab (Steps 1–2).
- Export a Hugging Face token if your model requires it:
export HF_TOKEN=your_actual_token_here
- Launch the recommended recipe for your hardware platform (Spark: Agent-ready Models; Station: Station deployment guide), ensuring
--host 0.0.0.0and port8000. - Confirm the API is up:
curl -s http://0.0.0.0:8000/v1/models
Expected: a JSON "data" array that includes your model handle. If the request hangs, the model is still loading — wait for Application startup complete in the container logs and retry.
- Note the exact model
idfrom the response. You will use that string foropenshell inference setand in the OpenClaw onboarding wizard.
OpenShell-specific requirements
| Requirement | Why it matters |
|---|---|
--host 0.0.0.0 | The OpenShell gateway runs inside Docker and cannot reach a server bound only to localhost. |
| Host IP in the provider URL | Create the OpenShell provider with http://{Machine_IP}:8000/v1, not localhost or 127.0.0.1. |
| Matching model id | The OpenClaw wizard Model ID must match the handle served by vLLM and configured in openshell inference set. |
Next steps
Return to the Instructions tab starting at Step 5. Serve a model with vLLM (or skip ahead if the server is already running), then create the inference provider and deploy the sandbox.