How I Installed OpenClaw Locally for Free

Search “install OpenClaw” on YouTube. The first 30 results are the same thing: a paid VPS, an affiliate link, and a tutorial copied from the official README. Nobody explains how to install it on a PC you already have at home.

I did it. An old PC with Ubuntu, a 6GB GPU, and zero euros per month. Here’s everything you need to know.

Why local AI

Three reasons:

1. Cost: 0 EUR/month. No subscription. No API bill. Open source models run on your hardware.
2. Privacy. Your prompts never leave your local network. No provider stores them, trains on them, or sells them.
3. Always on. An agent running 24/7 at home can automate tasks while you sleep. It doesn’t depend on having a browser tab open.

There’s a fourth reason that doesn’t get mentioned enough: with a residential IP you can scrape portals that block datacenter IPs. VPS ranges from Hetzner, DigitalOcean or AWS are blacklisted. Your home connection is not.

Hardware: what you actually need

This is what I use:

A few important notes:

• The GPU is not mandatory. Ollama can run models on CPU and RAM alone. But it’s 3-5x slower.
• 6 GB VRAM is a real limit. Models like Gemma 3 4B (3.3 GB) fit comfortably. Qwen 3.5 9B in Q4_K_M (~5.7 GB) barely fits. Anything above 9B needs more VRAM or offloads to CPU (slow).
• 32 GB RAM is the sweet spot. With 16 GB you can run models up to ~12B on CPU, but it gets tight if you want OpenClaw + Ollama + a browser open at the same time.
• Disk doesn’t matter if you use API. If you rely on OpenRouter you don’t need to download models (the largest ones are 20+ GB). You only need space if going full local.

My PC is from 2014. Literally a 12-year-old processor. If yours is newer, even better.

Step 1: Ubuntu 24.04

If you already have Linux installed, skip to step 2. If you’re coming from Windows, the safest route is dual boot: install Ubuntu on a separate partition without touching Windows.

I won’t cover Ubuntu installation here because there are 10,000 tutorials and every case is different (UEFI vs Legacy, SSD vs HDD, existing partitions). The only things that matter:

• Ubuntu 24.04 LTS (supported until 2029)
• Dedicated partition of at least 50 GB
• NVIDIA drivers installed (Ubuntu detects them automatically during installation, but verify)

To confirm the GPU is detected:

nvidia-smi

You should see something like:

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.xxx       Driver Version: 550.xxx       CUDA Version: 12.x              |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce GTX 980 Ti     Off  | 00000000:01:00.0  Off |                  N/A |
| 28%   34C    P8              16W / 250W |      0MiB /  6144MiB   |      0%      Default |
+-----------------------------------------+------------------------+----------------------+

If nvidia-smi doesn’t work, install the drivers:

sudo apt update
sudo ubuntu-drivers install
sudo reboot

Step 2: Install Ollama

Ollama is the runtime that executes AI models on your machine. Think Docker, but for LLMs.

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

Verify it’s running:

ollama --version

ollama version 0.21.0

Ollama installs as a systemd service and starts automatically. The server listens on localhost:11434.

Step 3: Download local models

This is where your VRAM matters. These are the models I’ve tested that fit in 6 GB:

Gemma 3 4B (recommended to start)

ollama pull gemma3:4b

Size: ~3.3 GB. Leaves VRAM headroom. On my GTX 980 Ti:

ollama run gemma3:4b "Explain what a reverse proxy is in 3 lines"

Speed: ~44 tokens/second on GPU. Fast. Enough for simple tasks: summaries, formatting, classification, short drafts.

Qwen 3.5 9B Q4_K_M (for higher quality)

ollama pull qwen3.5:9b-q4_K_M

Size: ~5.7 GB. Just barely fits in 6 GB VRAM. Slower than Gemma but noticeably smarter. Good for research, document analysis, long text generation.

Estimated speed: ~18 tokens/second on similar hardware.

Models that DON’T fit in 6 GB

• Qwen 3.6 235B — needs ~120 GB. Impossible locally. But it’s free on OpenRouter (next section).
• Nemotron 120B — needs ~60 GB+. Only viable in cloud.
• Qwen 3.5 27B — needs ~16 GB VRAM or ~32 GB RAM on CPU. Works on CPU with 32 GB RAM but it’s slow (~5-8 tok/s).

The reality: models you can run on 6 GB VRAM are good for simple tasks, but for complex reasoning you need something 200B+. That’s where OpenRouter comes in.

Step 4: Install OpenClaw

OpenClaw needs Node.js 22+ and git:

# Install Node.js 22 via NodeSource
curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt install -y nodejs git

# Verify
node --version  # v22.22.2 or higher
git --version

Install OpenClaw:

npm install -g @openclaw/cli

Verify the version:

openclaw --version

openclaw v2026.4.15

Step 5: Configure OpenClaw with Ollama (local models)

The first time you run OpenClaw it walks you through a setup wizard. But you can configure it manually:

openclaw configure

Select:

1. Provider: Ollama
2. Endpoint: http://localhost:11434 (default)
3. Model: gemma3:4b (or whichever you downloaded)

This creates the configuration at ~/.openclaw/openclaw.json.

Now start the gateway:

openclaw gateway start --port 18789 --bind 127.0.0.1

The --bind 127.0.0.1 flag is important: it only accepts local connections. If you expose it on 0.0.0.0 without authentication, anyone on your network can use your instance.

To make it persistent (start automatically when the PC boots), create a systemd service:

mkdir -p ~/.config/systemd/user/

cat > ~/.config/systemd/user/openclaw-gateway.service << 'EOF'
[Unit]
Description=OpenClaw Gateway
After=network.target ollama.service

[Service]
ExecStart=/usr/bin/openclaw gateway start --port 18789 --bind 127.0.0.1
Restart=on-failure
RestartSec=5
Environment=NODE_ENV=production

[Install]
WantedBy=default.target
EOF

systemctl --user daemon-reload
systemctl --user enable --now openclaw-gateway.service

Verify it’s running:

systemctl --user status openclaw-gateway.service

● openclaw-gateway.service - OpenClaw Gateway
     Loaded: loaded (~/.config/systemd/user/openclaw-gateway.service; enabled)
     Active: active (running) since ...

Step 6: Configure OpenRouter (for large models)

Local 4-9B models are fine for many things, but there are tasks where you need a 200B+ parameter model: complex analysis, long code generation, multi-step reasoning.

OpenRouter is an API gateway that gives you access to dozens of models. Some are free, including Qwen 3.6 (235B parameters) during its preview period.

Create account and get API key

1. Go to openrouter.ai
2. Create an account (free)
3. Go to Keys -> Create Key
4. Copy your key. The format is: sk-or-v1-...

Never share or publish your API key. Store it securely.

Configure in OpenClaw

openclaw configure

Select:

1. Provider: OpenRouter
2. API Key: paste your sk-or-v1-... key
3. Model: openrouter/qwen/qwen3-235b-a22b

This configures Qwen 3.6 (235B parameters, mixture of experts with 22B active) as your main model via API.

Restart the gateway:

openclaw gateway restart

Alternative: keep both providers

The ideal configuration uses Ollama for fast, cheap tasks (always free) and OpenRouter for tasks that need more power. You can switch between providers by editing ~/.openclaw/openclaw.json:

{
  "models": {
    "providers": {
      "ollama": {
        "endpoint": "http://localhost:11434"
      },
      "openrouter": {
        "apiKey": "sk-or-v1-..."
      }
    },
    "default": "openrouter/qwen/qwen3-235b-a22b"
  }
}

The default field determines which model OpenClaw uses by default. Change it to ollama/gemma3:4b when you want zero cost.

Step 7: Test that everything works

Basic gateway test

curl http://localhost:18789/health

{"status": "ok", "version": "2026.4.15"}

Chat test with the model

openclaw chat "What version of OpenClaw am I running?"

If it responds coherently, everything is working. If you get a connection error, check that the gateway is active (systemctl --user status openclaw-gateway.service).

Test Ollama directly

curl http://localhost:11434/api/generate -d '{
  "model": "gemma3:4b",
  "prompt": "Hello, respond in one line",
  "stream": false
}'

It should return a JSON with the model’s response.

What I use it for

I didn’t install this to play around. I use it to automate real work:

• Automated scraping. I have cron jobs that launch scrapers every 6 hours and OpenClaw processes the data: cleans duplicates, classifies by brand, calculates prices.
• Research. I feed it PDFs, articles or technical documentation and it returns structured summaries.
• Drafts. Product descriptions, landing page copy, blog posts (not this one — I wrote this one myself).
• Data analysis. I give it a CSV with thousands of rows and ask it to find anomalies or patterns.

The differentiator versus using ChatGPT in the browser: this runs unattended. I can schedule tasks at 3AM and review the results in the morning. It’s an agent, not a chatbot.

Cost breakdown

When the Qwen 3.6 free preview on OpenRouter ends, the per-token cost will be minimal — we’re talking cents per conversation. And you always have local models as a free fallback.

Compare that to the alternatives:

• Claude/ChatGPT subscription: 20 EUR/month (and you can’t use them as an autonomous agent)
• VPS with GPU (Lambda, Vast.ai): 50-200 EUR/month
• Anthropic or OpenAI API without limits: variable, but easily 30+ EUR/month with moderate usage

Limitations (being honest)

• 6 GB VRAM limits you to small models. Gemma 3 4B and Qwen 3.5 9B are useful but don’t compete with GPT-4 or Claude Opus on complex tasks. For that you need OpenRouter.
• A 2014 i5 is not fast. CPU inference is viable but slow. If you plan to run 27B+ models on CPU, be patient.
• No Docker, no sandbox. OpenClaw has a sandbox mode based on Docker. I don’t have it installed, so commands the agent executes have full system access. Be careful with what you ask it to do.
• Initial setup is not trivial. If you’ve never touched Linux, installing Ubuntu + NVIDIA drivers + Ollama + Node.js + OpenClaw can take an afternoon. But you do it once.

Conclusion

All OpenClaw content on the internet assumes you’re going to pay for a VPS. Nobody talks about the most obvious option: use a PC you already have.

A 2014 computer with Ubuntu, 32 GB of RAM and a 6 GB GPU is enough to have your own AI agent running 24/7. Free. Without depending on any provider.

The complete setup took me about 3 hours including the Ubuntu installation. If you already have Linux, you can have it running in an hour.

If you have questions or want to see how I use it to automate motorcycle price scraping, subscribe — more articles are on the way.

Updated: April 22, 2026. OpenClaw v2026.4.15, Ollama v0.21.0, Ubuntu 24.04 LTS.