Skip to main content
Last Updated: March 31, 2026

​
Introduction

NVIDIA NemoClaw is an open-source reference stack from that runs OpenClaw inside a sandboxed container environment NVIDIA OpenShell. It adds OS-level filesystem and network isolation to OpenClaw agents, making it safer to run autonomous AI assistants on your machine. NemoClaw supports custom OpenAI-compatible inference endpoints, so you can point it at a self-hosted model on SaladCloud instead of a cloud provider. NemoClaw pairs well with SaladCloud because:
  • OpenAI-compatible endpoint support - the onboarding wizard has a custom option for any compatible inference server
  • Per-hour pricing - no per-token costs for long-running assistant sessions
  • Sandboxed execution - NemoClaw’s container isolation limits what your AI agent can do on your local machine
NemoClaw is currently in alpha (as of March 2026). APIs, configuration schemas, and runtime behavior are subject to breaking changes between releases. It is not recommended for production use.

​
Prerequisites

Before getting started, make sure you have:
  • A SaladCloud account
  • Docker installed (Linux/Windows WSL2) or Apple Container (macOS)
  • A Linux, macOS, or Windows (WSL2) machine

​
Step-by-Step Setup

​
Step 1: Deploy an LLM Recipe on SaladCloud

First, deploy an OpenAI-compatible LLM server on SaladCloud.
  • Go to the SaladCloud portal and create an account if you do not already have one.
  • Create an organization or choose an existing one, then click “Deploy a container group”.
  • Select an LLM recipe. The Qwen3.5-35B-A3B (llama.cpp) recipe is well-suited for conversational assistant use cases. On the recipe page, provide a name and deploy - the rest is preconfigured with recommended settings.
  • Once deployed, your endpoint will be live and serving an OpenAI-compatible API.
Available recipes: Ready-to-deploy recipes (best for less technical users):
  • qwen3.5-35B-A3B - A powerful Mixture of Experts model optimized for instruction-following tasks, ideal for agentic use cases.
  • qwen3.5-9b-llama-cpp - Optimized for Qwen3.5 9B model.
Recipes for custom deployments (best for advanced users):
  • llama.cpp - Supports GGUF models
  • sglang - High-performance inference
  • vllm - Popular LLM serving framework
  • ollama - Simple model management
  • tgi - Hugging Face Text Generation Inference server
After deployment, note your:
  • API endpoint URL (e.g., https://your-endpoint.salad.cloud)

​
Step 2: Install NemoClaw

Run the official install script: 
curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
Verify the installation: 
nemoclaw --version

​
Step 3: SaladCloud Authentication

For easy setup, when deploying your SaladCloud container group, leave the Authentication toggle off. Your endpoint will be accessible without a key. If you prefer to keep authentication enabled, you can still use NemoClaw by providing a dummy API key during onboarding and hardcoding your actual key in the config file after setup (see Step 4). You need to connect to the sandbox and update config file : /sandbox/.openclaw/openclaw.json. Add the following headers section under models.providers 
"headers": {
  "Salad-Api-Key": "your-actual-api-key"
}

​
Step 4: Run the Onboarding Wizard

Start NemoClaw onboarding: 
nemoclaw onboard
When prompted to select an inference endpoint:
  1. Choose Other OpenAI-compatible endpoint from the provider list
  2. Enter your SaladCloud endpoint URL when prompted: https://your-endpoint.salad.cloud/v1
  3. For OpenAI-compatible endpoint API key, enter any non-empty string (e.g., dummy) since authentication is disabled
  4. For the model name, enter the name of the model you deployed (e.g., qwen3.5-35b-a3b)
  5. NemoClaw will build a container and validate connectivity to the endpoint before completing setup
  6. Follow the remaining prompts to finish configuration

​
Step 5: Test the Connection

Once onboarding completes you can connect to your sandboxed assistant by running nemoclaw <your_assistant_name> connect and then connecting to it via the local TUI (openclaw tui), by accessing local dashboard (ws://127.0.0.1:18789) or by setting up a messaging channel (e.g., Telegram bot).
“Hello! Summarize what SaladCloud is in two sentences.”
If NemoClaw responds, your setup is complete.

​
Model Recommendations

  • Qwen 3.5-35B-A3B: Best for complex assistant tasks and agentic workflows
  • Qwen 3.5-9B: Suitable for simple Q&A and quick responses