01 — Overview and Stack

This tutorial walks you through deploying OpenClaw (a self-hosted AI agent platform) on AMD Developer Cloud, backed by an open-source Large Language Model (LLM) served via vLLM on AMD's Radeon Open Compute (ROCm) software stack.

The setup enables you to run OpenClaw as a persistent cloud agent for tasks like model evaluation, file operations, and web-search-augmented research, without tying up your local hardware.

Component	Value
GPU	AMD MI300X, 192 GB Video RAM (VRAM)
Platform	AMD Developer Cloud (DigitalOcean GPU Droplet)
Model	Qwen/Qwen3-30B-A3B — Mixture-of-Experts (MoE), 30B total / 3B active per token, text only, served as qwen3-30b
Inference server	vLLM on ROCm, Docker image vllm/vllm-openai-rocm:latest
Tool call parser	hermes — confirmed valid in this ROCm vLLM build
Web search	SearXNG, self-hosted Docker container, no Application Programming Interface (API) key required
vLLM port	8000
SearXNG port	8888
OpenClaw Control UI port	18789
Setup method	Manual Secure Shell (SSH) from local Windows machine

---

02 — Prepare Your Local Machine

Complete these steps on the command prompt of your local Windows machine before creating the cloud droplet.

Verify SSH is available on your machine:

powershell.exe -NoProfile -Command "ssh -V"

Check whether an SSH key already exists:

type %USERPROFILE%\.ssh\id_ed25519.pub

An error message The system cannot find the path specified indicates no SSH key exists. Create a new SSH key with the following command:

ssh-keygen -t ed25519 -C "amd-droplet"

vLLM uses your HuggingFace (HF) token to download model weights during startup.

1. Log in at huggingface.co.
2. Go to Settings → Access Tokens.
3. Click Create new token and select Read permission.
4. Copy the token value. You will paste it on the droplet in Step 5.

---

03 — Create the GPU Droplet

A GPU Droplet is a cloud virtual machine (VM) with a dedicated Graphics Processing Unit. AMD Developer Cloud provides access to MI300X GPU Droplets via DigitalOcean.

Log in to cloud.digitalocean.com and follow these steps:

1. Click Create → Droplets.
2. Hardware: select the single MI300X instance (gpu-mi300x1-192gb).
3. Image: select ROCm Software.
   Why ROCm: ROCm is AMD's open-source GPU compute platform. It includes the drivers and libraries required to run vLLM on AMD hardware. A generic Ubuntu image would require manual ROCm installation.
4. Authentication: click Add SSH Key and select or paste your public key from Step 1.
5. Click Create Droplet and wait 30–60 seconds for provisioning to complete.
6. Copy the public Internet Protocol version 4 (IPv4) address shown on the overview page — you will use it in every SSH command that follows.

---

04 — Set Up the Droplet

Run all commands in this section inside the droplet via SSH, in the same terminal window unless stated otherwise.

Open Command Prompt on your local machine and run:

ssh root@<YOUR_DROPLET_IPV4_ADDRESS>

Write your tokens to the droplet's shell environment. Replace the placeholder values with your actual tokens. These variables persist across reconnections because they are written to .bashrc.

cat >> /root/.bashrc << 'EOF'
export HF_TOKEN=<your_huggingface_read_token>
export MY_API_KEY=<your_chosen_secret_key>
EOF
source /root/.bashrc

Verify both variables are set:

echo $HF_TOKEN
echo $MY_API_KEY

OpenClaw requires Node.js version 24 or later.

curl -fsSL https://deb.nodesource.com/setup_24.x | bash -
apt-get install -y nodejs

Verify Node.js installed correctly:

node --version

Install OpenClaw globally:

npm install -g openclaw@latest

vLLM is a high-throughput inference engine. This command starts vLLM in a Docker container, downloads the Qwen3 model weights from HuggingFace (~60 GB on first run), and exposes a standard chat completions API on port 8000.

docker run -d \
  --name vllm-qwen3 \
  --ipc=host --network=host --privileged \
  --cap-add=CAP_SYS_ADMIN --cap-add=SYS_PTRACE \
  --device=/dev/kfd --device=/dev/dri \
  --group-add=video \
  --security-opt=seccomp=unconfined \
  --shm-size 16G \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -e HF_TOKEN=$HF_TOKEN \
  vllm/vllm-openai-rocm:latest \
  Qwen/Qwen3-30B-A3B \
  --served-model-name qwen3-30b \
  --api-key $MY_API_KEY \
  --host 0.0.0.0 --port 8000 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.80 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes

docker logs -f vllm-qwen3

curl http://127.0.0.1:8000/v1/models \
  -H "Authorization: Bearer $MY_API_KEY"

curl -s http://127.0.0.1:8000/v1/chat/completions \
  -H "Authorization: Bearer $MY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3-30b","messages":[{"role":"user","content":"Say hello"}],"max_tokens":50}' \
  | python3 -m json.tool

SearXNG is a self-hosted, open-source meta-search engine that aggregates results from Google, Bing, DuckDuckGo and others, with no API key required. OpenClaw uses it to give the agent live web access.

Start a SearXNG container on the droplet:

docker run -d -p 8888:8080 --name searxng searxng/searxng

Run in your SSH session. The token is generated randomly and stored as a shell variable for the current session. Do not close this terminal before Step 12.

GATEWAY_TOKEN=$(openssl rand -hex 24)
echo "Gateway token: $GATEWAY_TOKEN"
mkdir -p ~/.openclaw

Run immediately after Step 9 in the same terminal. The shell substitutes $GATEWAY_TOKEN and $MY_API_KEY automatically. Do not edit the values manually.

cat > ~/.openclaw/openclaw.json << CONFIGEOF
{
  "gateway": {
    "mode": "local",
    "auth": { "mode": "token", "token": "${GATEWAY_TOKEN}" },
    "controlUi": { "allowInsecureAuth": true }
  },
  "models": {
    "mode": "merge",
    "providers": {
      "vllm": {
        "baseUrl": "http://127.0.0.1:8000/v1",
        "apiKey": "${MY_API_KEY}",
        "api": "openai-completions",
        "models": [{
          "id": "qwen3-30b", "name": "Qwen3 30B",
          "reasoning": false, "input": ["text"],
          "cost": {"input":0,"output":0,"cacheRead":0,"cacheWrite":0},
          "contextWindow": 32768, "maxTokens": 4096
        }]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": { "primary": "vllm/qwen3-30b" },
      "sandbox": { "mode": "all" }
    }
  },
  "plugins": { "entries": { "vllm": { "enabled": true } } },
  "tools": {
    "web": {
      "search": { "enabled": true, "provider": "searxng" },
      "searxng": { "baseUrl": "http://127.0.0.1:8888" }
    }
  }
}
CONFIGEOF

Verify the config saved with the correct values:

cat ~/.openclaw/openclaw.json | grep -E "primary|baseUrl|id|provider"

openclaw onboard

When prompted, enter these values:

Model provider:  vLLM
Base URL:        http://127.0.0.1:8000/v1
API Key:         <your MY_API_KEY value>
Model ID:        qwen3-30b
Channel:         Skip for now

openclaw gateway install

openclaw config set gateway.mode local
openclaw gateway start
openclaw status

openclaw dashboard

---

05 — Access OpenClaw from Your Local Machine

The OpenClaw gateway binds only to the loopback interface (127.0.0.1) on the droplet for security. An SSH tunnel forwards the port to your local machine.

Run this on your local Windows machine, not inside the droplet. Open a new Command Prompt window, keeping your droplet SSH session open separately.

ssh -L 18789:127.0.0.1:18789 -N root@<YOUR_DROPLET_IPV4_ADDRESS>

Paste the tokenised URL from Step 13 into your browser. Do not use http://127.0.0.1:18789 directly. Use the full tokenised URL.

1. Click the Chat tab.
2. Select vllm/qwen3-30b from the model dropdown.
3. Type /new and press Send to start a fresh session.
4. Type: Tell me your name and press Send.

---

06 — Troubleshooting

vLLM Startup Errors

Error	Fix
KeyError: invalid tool call parser	Only hermes is valid for Qwen3 in this build. Do not use llama, llama3_json, gemma4, or pythonic.
Connection refused on port 8000	vLLM is still loading. Wait for Application startup complete in docker logs.
{"error":"Unauthorized"}	Wrong API key. Run: echo $MY_API_KEY.
Container exits immediately	Run docker logs vllm-qwen3 to read the error before the container exits.

Gateway Errors

Error	Fix
Gateway: unreachable after gateway start	Run openclaw gateway install then openclaw gateway start.
disconnected (1006) in browser	SSH tunnel dropped. Re-open in a new local CMD window: ssh -L 18789:127.0.0.1:18789 -N root@<YOUR_DROPLET_IP>
Permission denied (publickey) on tunnel	Run the tunnel command on the command prompt of your local Windows machine, not on the droplet.
openclaw logs --follow: pairing required	Use: tail -f /tmp/openclaw/openclaw-$(date +%Y-%m-%d).log

Restart vLLM

docker stop vllm-qwen3 && docker rm vllm-qwen3
# Then re-run the full docker run command from Step 4.

---

07 — Shut Down and Destroy the Droplet

Always destroy the droplet at the end of each session; powering it off does not stop billing.

Stop Services on the Droplet

openclaw gateway stop
docker stop vllm-qwen3
docker stop searxng

Destroy on DigitalOcean

1. Go to cloud.digitalocean.com and select your droplet.
2. Click Destroy in the left sidebar.
3. Click Destroy this Droplet and confirm.

08 — Known Issues and Caveats

When selecting a model to use with OpenClaw on vLLM, keep these compatibility constraints in mind. The issues below are specific to OpenClaw's request format, the vLLM ROCm build, and Anthropic's authentication policy — not limitations of AMD Developer Cloud or the MI300X hardware.

---