OpenClaw on Jetson

🦞 OpenClaw is a personal digital assistant that runs directly on your device. Unlike typical chatbots, OpenClaw can actively manage files, browse the web, and control applications, integrating with messaging platforms like WhatsApp, Discord, and Telegram to perform tasks on your behalf. It memorizes who you are, your patterns, and your preferences over time, and you can give it all sorts of tools to help automate parts of your life.

Looking for inspiration? Check out what people have built with OpenClaw at openclaw.ai/showcase. It’ll give you a good idea of what’s possible.

In this tutorial, we’ll walk you through setting up OpenClaw with everything running 100% locally on a Jetson, no cloud APIs needed. All you need is a Jetson and a phone with WhatsApp. By the end, you’ll have a fully working AI agent that you can talk to through your phone, and it can actually do things for you.

Prerequisites

Step 1: Serve a Local Model with vLLM

Before setting up OpenClaw, we need to host a model locally. For this tutorial we’ll use vLLM as the serving engine.

Any model should work here as long as it’s capable of tool calling. Tool calling is very important for OpenClaw. It’s how the agent takes actions on your behalf.

Tip: In our testing, we found Mixture of Experts (MoE) models work exceptionally well with OpenClaw, models like Nemotron 3 Nano 30B-A3B, Qwen 3.5 35B-A3B, and GLM 4.7 Flash. That said, any model that supports tool calling should work. Just make sure you serve it correctly with vLLM so it can handle tool use without messing up the output format.

Export your Hugging Face token

Some models require you to accept a license agreement on Hugging Face before using them. Export your token so vLLM can download the model:

export HF_TOKEN=your_huggingface_token_here

Serve the model

For this tutorial, we’ll go with Nemotron 3 Nano 30B-A3B. Select your device below:

sudo docker run -it --rm --pull always \
  --runtime=nvidia --network host \
  -e HF_TOKEN=$HF_TOKEN \
  -e VLLM_USE_FLASHINFER_MOE_FP4=1 \
  -e VLLM_FLASHINFER_MOE_BACKEND=throughput \
  -v $HOME/.cache/huggingface:/data/models/huggingface \
  ghcr.io/nvidia-ai-iot/vllm:latest-jetson-thor \
  bash -c "wget -q -O /tmp/nano_v3_reasoning_parser.py \
  --header=\"Authorization: Bearer \$HF_TOKEN\" \
  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/resolve/main/nano_v3_reasoning_parser.py \
  && vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 \
  --gpu-memory-utilization 0.8 \
  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser-plugin /tmp/nano_v3_reasoning_parser.py \
  --reasoning-parser nano_v3 \
  --kv-cache-dtype fp8"

sudo docker run -it --rm --pull always \
  --runtime=nvidia --network host \
  -e HF_TOKEN=$HF_TOKEN \
  -e VLLM_USE_FLASHINFER_MOE_FP4=1 \
  -e VLLM_FLASHINFER_MOE_BACKEND=throughput \
  -v $HOME/.cache/huggingface:/data/models/huggingface \
  ghcr.io/nvidia-ai-iot/vllm:latest-jetson-orin \
  bash -c "wget -q -O /tmp/nano_v3_reasoning_parser.py \
  --header=\"Authorization: Bearer \$HF_TOKEN\" \
  https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4/resolve/main/nano_v3_reasoning_parser.py \
  && vllm serve stelterlab/NVIDIA-Nemotron-3-Nano-30B-A3B-AWQ \
  --gpu-memory-utilization 0.8 \
  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser-plugin /tmp/nano_v3_reasoning_parser.py \
  --reasoning-parser nano_v3 \
  --kv-cache-dtype fp8"

Tip: These models need a lot of memory. Before serving, make sure you don’t have other processes eating up GPU memory. It’s always a good idea to clear the memory cache first to ensure you have as much free memory as possible:

sudo sysctl -w vm.drop_caches=3

Note: You will need an AGX Thor or AGX Orin to run the models mentioned above. If you’re on a different Jetson (like Orin Nano Super), any smaller model should work fine as long as it supports tool calling, however it may not be as capable as these larger models.

Verify the model is serving

Once vLLM is up and running, open a separate terminal and verify:

curl -s http://127.0.0.1:8000/v1/models

You should see your model listed in the response. Once you see it, you’re ready to move on.

Step 2: Install Node.js 22+

OpenClaw requires Node.js 22 or newer. Let’s get that installed:

curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt install -y nodejs

Quick sanity check:

node --version   # should be v22.x.x or higher

Step 3: Install OpenClaw

Now let’s install OpenClaw itself:

sudo npm install -g openclaw@latest

And verify it’s there:

openclaw --version

Step 4: Run the Onboarding Wizard

This is where the fun begins! OpenClaw has an interactive wizard that sets up everything for you: model provider, gateway, WhatsApp, workspace, and hooks.

openclaw onboard --skip-daemon

Why --skip-daemon? The systemd daemon installer has a known issue on headless/SSH sessions. We’ll skip it here and start the gateway manually in Step 5 instead.

The wizard will walk you through several steps. Here’s what to expect:

Model / Auth provider

One of the first things you’ll be asked is to select a model provider. We’re linking OpenClaw to our local Nemotron 3 Nano through vLLM. Select vLLM and then configure:

Channel selection (WhatsApp)

There are different ways to talk to your OpenClaw agent. The simplest is through WhatsApp, which is what we’ll be using in this tutorial.

Select WhatsApp as the channel. A QR code will show up in your terminal. Grab your phone:

1. Open WhatsApp > Settings > Linked Devices
2. Tap Link a Device
3. Scan the QR code

Give it a few seconds and you should see:

Linked after restart; web session ready.

It will then ask for your phone number. Enter your WhatsApp phone number in international format (e.g. +15551234567).

Skills

Next up is skills. These are all the skills that can be installed through the onboarding wizard. If any appeal to you, feel free to install them. For this tutorial, we’ll be skipping all of them. Don’t worry, you can very easily add any skill later.

API keys for cloud services

The next step will ask you for a bunch of API keys for different services (Brave Search, Perplexity, etc.). Since we’re going fully local, we can safely say no to all of them. If you want to use any of these cloud services later, you can freely add your API key and it will be registered for use.

Hooks

Hooks let you automate actions when agent commands are issued. For example, saving session context to memory when you issue /new or /reset. Feel free to select or deselect whichever you like. For this tutorial we’ll be selecting all of them since they can be very useful.

Bot hatching

Finally, when it asks “How do you want to hatch your bot?”, feel free to say “I’ll do this later” since we’ll be interacting with our bot through WhatsApp.

And voilà, OpenClaw is installed and configured!

Step 5: Start the Gateway

Now we need to get the gateway running. In the same terminal, run:

nohup openclaw gateway run > /tmp/openclaw-gateway.log 2>&1 &

What’s happening here? This runs the gateway in the background and logs everything to /tmp/openclaw-gateway.log in case we need it for debugging later.

Give it a few seconds to start up, then check the status:

openclaw channels status --probe

You should see something like this:

Gateway reachable.
- WhatsApp default: enabled, configured, linked, running, connected, dm:pairing

If you see that, everything is working and you’re good to go!

Step 6: Talk to Your Agent Through WhatsApp

This is the moment of truth! OpenClaw links as a secondary device to your WhatsApp account (similar to WhatsApp Web). From here, you can start communicating with your agent by opening your own chat in WhatsApp (“Message yourself”). It should reply back, and you should see the messages right there in the conversation.

Tip: The first message may take a moment as the model warms up. Don’t worry, subsequent messages will be much faster.

Here’s an example of what it looks like:

That’s it. You now have a fully local AI agent running on your Jetson, accessible through WhatsApp, with zero cloud dependencies. Pretty cool, right?

The sky is the limit from here. You can ask it to search things for you, have it watch if a price drops and message you when it happens, or automate parts of your daily workflow. Some people have even jokingly had OpenClaw join conversations to talk as them. Really, the sky is the limit.

Useful WhatsApp Commands

Once you’re chatting with your agent, these built-in commands can come in handy. They work directly in the chat without invoking the LLM:

Managing the Gateway

Here are some handy commands for managing the gateway day-to-day:

# Start the gateway
nohup openclaw gateway run > /tmp/openclaw-gateway.log 2>&1 &

# Stop the gateway
pkill -f "openclaw gateway run"

# Restart (stop + start)
pkill -f "openclaw gateway run"; sleep 2
nohup openclaw gateway run > /tmp/openclaw-gateway.log 2>&1 &

# View logs (live)
tail -f /tmp/openclaw-gateway.log

# Or use OpenClaw's built-in log viewer (while gateway is running)
openclaw logs --follow

# Check status
openclaw channels status --probe

Troubleshooting

If you run into issues, here are the most common ones and how to fix them:

Unlink WhatsApp (Cleanup)

If you ever want to remove the OpenClaw link from your WhatsApp account:

On your phone: WhatsApp > Settings > Linked Devices → tap the OpenClaw session → Log Out.

What to Do Next

• Explore OpenClaw’s showcase for ideas on what to build
• Try installing skills to give your agent new capabilities
• Experiment with different models. Swap in any tool-calling model and see how it performs
• Check out the OpenClaw website for documentation and community resources