OpenClaw on SwissGPU

Introduction

OpenClaw is a self-hosted gateway for AI agents. It runs as a single gateway process on your server and provides a browser control UI for chat, sessions, configuration, and model setup.

It can be used directly in the browser, or connected to channels such as WhatsApp, Telegram, Discord, and iMessage. The main benefit is control: the gateway, workspace, and model connection stay on your own machine instead of depending on a hosted service.

On a SwissGPU server, OpenClaw works well as the control layer in front of local GPU-backed model runtimes such as Ollama or vLLM. The usual setup is to install OpenClaw, complete the onboarding wizard, confirm that the gateway is running, and then connect a local model provider on the same machine.

What You Need

An active SwissGPU Linux server with an NVIDIA GPU.
Working SSH access.
A user account with sudo privileges.
Enough disk space for OpenClaw and at least one local model.

Step 1: Verify GPU Availability

After logging in by SSH, check that the GPU is visible:

nvidia-smi

Step 2: Install OpenClaw

curl -fsSL https://openclaw.ai/install.sh | bash

The installer sets up OpenClaw and may ask forsudoso it can install Node.js and the base packages OpenClaw needs.

Step 3: Complete the Onboarding Wizard

After installation, OpenClaw opens its onboarding wizard automatically. The wizard is the simplest way to accept the security notice, choose a local gateway, and configure a first model provider.

If you close it or want to run it again later, start it with:

openclaw onboard --install-daemon

For a local setup, chooseLocal (this machine)when OpenClaw asks where the gateway should run.

Step 4: Put OpenClaw on PATH

To make the OpenClaw command available in new shells, add the npm user bin directory to your shell path:

export PATH="$HOME/.npm-global/bin:$PATH"
echo 'export PATH="$HOME/.npm-global/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc

If your server useszshinstead ofbash, write the same export line to~/.zshrcinstead.

Step 5: Confirm the Local Gateway

systemctl --user enable --now openclaw-gateway.service
openclaw gateway status
openclaw health
curl -I http://127.0.0.1:18789/

Once the gateway is running, the local dashboard is available on127.0.0.1:18789.

If you want OpenClaw to keep running after logout and reboot, enable linger once:

sudo loginctl enable-linger "$USER"

Step 6: Manual Gateway Setup

If the wizard did not create the local gateway, or if you want to recreate it manually, run:

openclaw gateway install
openclaw config set gateway.mode '"local"'
openclaw config set gateway.bind '"loopback"'
openclaw config validate

You can print the gateway token later with:

openclaw config get gateway.auth.token

Step 7: Optional Browser Access

If you want to open the OpenClaw UI in a browser on the local machine, forward the local gateway port:

ssh -N -L 18789:127.0.0.1:18789 <your-existing-ssh-target>

Once the tunnel is open, the local browser can usehttp://127.0.0.1:18789/.

Step 8: Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

The official Linux installer sets up the Ollama binary and systemd service.

Step 9: Pull and Test an Ollama Model

Check that the service is active:

sudo systemctl status --no-pager ollama
ollama --version

Download a small model first:

ollama pull llama3.2:1b

Then test the local API and confirm the model is using the GPU:

cat >/tmp/ollama-test.json <<'EOF'
{
  "model": "llama3.2:1b",
  "prompt": "Reply only: OK",
  "stream": false,
  "keep_alive": "5m"
}
EOF
curl -s http://127.0.0.1:11434/api/generate -d @/tmp/ollama-test.json
ollama ps

Step 10: Connect Ollama to OpenClaw

If the onboarding wizard already configured Ollama, you can skip this step.

Otherwise run the focused model configuration flow:

openclaw configure --section model

ChooseLocal (this machine)for the gateway location, then chooseOllamaas the provider.

Use these values when OpenClaw asks for the local provider details:

Provider: Ollama
Base URL: http://127.0.0.1:11434
API key: ollama-local
Model: llama3.2:1b

Step 11: Install vLLM Prerequisites

A clean Ubuntu server does not always include the packages needed forpython3 -m venvandpip. Install them first:

sudo apt-get update
sudo apt-get install -y python3-pip python3.12-venv

Step 12: Install and Start vLLM

Create an isolated environment and install vLLM:

python3 -m venv ~/vllm-test-env
source ~/vllm-test-env/bin/activate
python -m pip install --upgrade pip
python -m pip install vllm

If Ollama is still holding a model in GPU memory, unload it before starting vLLM on the same GPU:

ollama stop llama3.2:1b

Start the server:

vllm serve Qwen/Qwen2.5-0.5B-Instruct \
  --host 127.0.0.1 \
  --port 8000 \
  --api-key vllm-local \
  --gpu-memory-utilization 0.8

Once the model is loaded, confirm that the OpenAI-compatible endpoint is available:

curl -s http://127.0.0.1:8000/v1/models -H 'Authorization: Bearer vllm-local'

Step 13: Connect vLLM to OpenClaw

If the onboarding wizard already configured vLLM, you can skip this step.

Otherwise run the focused model configuration flow again:

openclaw configure --section model

ChooseLocal (this machine)for the gateway location, then choosevLLMas the provider.

Use these values for the vLLM provider:

Provider: vLLM
Base URL: http://127.0.0.1:8000/v1
API key: vllm-local
Model: Qwen/Qwen2.5-0.5B-Instruct

Troubleshooting

openclaw gateway status
journalctl --user -u openclaw-gateway.service -n 100 --no-pager
ollama ps
tail -n 120 /tmp/vllm.log
nvidia-smi

OpenClaw is not found in a new shell: add~/.npm-global/binto the shell path and reload the shell.
The gateway service starts and exits: confirm thatgateway.modeis set tolocalandgateway.bindis set toloopback.
The gateway probe fails right after restart: wait a few seconds and run the status command again.
Python venv creation fails: installpython3.12-venvfirst.
vLLM fails with a GPU memory error: stop the loaded Ollama model or lower--gpu-memory-utilization.
The vLLM models endpoint returns Unauthorized: pass the bearer token header in thecurlrequest.

SwissGPU