Start Your Local Openclaw Bot with 1 Click

This guide will illustrate how to deploy and use the OpenClaw-vLLM Docker image, which provides a complete AI agent environment with local inference.

🏁What's Included

• vLLM Server: High-performance inference with tool calling support

• OpenClaw Gateway: Web-based AI agent interface

• Qwen3-30B Model: Pre-loaded 30B parameter model with MoE architecture

🏁Requirements

• Recommended(default) setting of GPUs: 4 * RTX 5090

• Make sure you have at least 80GB of VRAM

If you use more/less GPUs, please alter here

• Docker or compatible container runtime

• Access to the YottaLabs container registry

1

Deploy the container

Launch the image on your GPU instance. The container automatically:

1. Starts vLLM server on port 8001

2. Configures OpenClaw with the local vLLM endpoint

3. Launches OpenClaw Gateway on port 18789

Startup should take a while as vLLM loads the model into GPU memory. Go grab a coffee ☕ !

2

SSH Tunnel Configuration

OpenClaw Gateway is configured in mode: "local", listening on localhost.

Basic SSH Tunnel

ssh -L 8001:localhost:8001 -L 18789:localhost:18789 \
    <USER>@<REMOTE_HOST> -p <SSH_PORT> -i <KEY_FILE>
e.g.ssh -L 8001:localhost:8001 -L 18789:localhost:18789 [email protected] -p 30059 -i private_key.pem

YottaLabs Instance Example

ssh -L 8001:localhost:8001 -L 18789:localhost:18789 \
    [email protected] -p 30061 -i ~/.ssh/yotta_key.pem

Parameter Substitution

PARAMETER	YOTTALABS EXAMPLE	CUSTOM VALUE
<USER>	user	Your SSH username
<REMOTE_HOST>	your-instance.yottalabs.ai	Pod hostname
<SSH_PORT>	30061	Assigned SSH port
<KEY_FILE>	~/.ssh/yotta_key.pem	Path to private key

3

Access Services

Visit URL: http://localhost:18789

Model Details

The included Qwen3-30B-A3B-Instruct model features:

• 30B parameters with Active MoE architecture

• 32K context window

• Native tool calling support

• Optimized for instruction following vLLM not starting

Check the logs:

tail -f /workspace/vllm.log