Dell Technologies Dell Deskside Agentic AI Solutions
Home / solutions / nemoclaw / landscape-fleet-deploy

Fleet deployment via Canonical Landscape (OpenClaw on NemoClaw)

Roll OpenClaw on NemoClaw out to a fleet of enrolled Ubuntu systems (Dell deskside systems powered by NVIDIA) from Canonical Landscape, with no SSH, logged in the Landscape activity feed and repeatable from CI. The enrolment, script-execution, verification and close-out steps are the same for every solution and are described once in Fleet Deploy via Landscape (Agentic Coding); this page adds what is specific here.

Steps

  1. Enable script execution on the target devices (Step 1 of the common guide, root, once per rollout window).
  2. Stage this repository on each device at /opt/deskside-agentic-ai-solutions-for-nvidia (file transfer only).
  3. Deploy with the solution payload:
export SOLUTION=openclaw
python3 scripts/landscape/landscape_client.py runscript \
  --file scripts/landscape/deploy-solution.sh --ids "$ID" --user root --limit 3000
python3 scripts/landscape/landscape_client.py pollscript --group <group-id> --timeout 3200

The payload runs the serving recipe first (nvidia runtime self-heal, pinned image, readiness contract), then the solution step, and exits non-zero unless the solution is actually up. Variables: NEMOCLAW_INSTALLER_SHA256 (required pin of NVIDIA's installer), AGENT_USER (account that owns the runtime, default dell; it is added to the docker group), LLM_MODEL (default: the first model the local endpoint serves), SANDBOX_NAME, DESKSIDE_API_KEY, NEMOCLAW_INSTALL_REF (pin the NemoClaw release). 4. Verify (Step 3 of the common guide, scripts/landscape/verify-serving.sh) and confirm openshell status is healthy on the device, the sandbox lists as Ready, and a prompt sent from inside the sandbox reaches the local endpoint at http://host.docker.internal:8000/v1. 5. Close root script execution (Step 4 of the common guide, scripts/landscape/disable-script-execution.sh) once the fleet is healthy.

Notes

  • Onboarding runs through NemoClaw's own non-interactive mode, pointed at the local OpenAI-compatible endpoint (NEMOCLAW_PROVIDER=custom), so no wizard is needed and no answers file has to be captured.
  • Plane 2 installs from NVIDIA's shipping installer. The payload fetches it to a file and refuses to run it unless its sha256 matches your pin.
  • Roll out in waves (one tagged canary first); sandboxes keep running while the serving container is replaced.

Rollback and teardown

Re-run the deploy with the previous image digest from recipes/versions.lock, or remove the containers with a one-line payload (docker rm -f <container>; docker compose down -v for compose-based solutions). Re-enable script execution first if it has been closed.