How to Run OpenClaw for Free and Offline with Ollama | OpenClaw DC

Yes, OpenClaw is 100% free and can run completely offline using Ollama with local models. You need no API key, no subscription, and no internet connection after the initial setup. The tradeoff is you need a machine with at least 16GB RAM for reliable tool-calling models. OpenClaw itself is open-source, and Ollama is a free local model server. Together they give you a fully private, zero-cost AI agent that runs on your own hardware. This guide covers every step from first install to your first offline task.

Step 1: Install Ollama

Ollama is a lightweight model server that runs LLMs locally. Download it from ollama.com or install from the command line.

macOS (Homebrew):

brew install ollama

Linux:

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

Windows: Download the installer from ollama.com/download. WSL2 is recommended if you plan to run OpenClaw on Windows.

Once installed, start the Ollama server:

ollama serve

Ollama runs in the background on port 11434 by default. Leave it running.

Step 2: Pull a Local Model

You need a model that supports tool calling. Not every model works well with OpenClaw. Our testing shows Qwen3.5 27B is the best balance of quality, speed, and RAM usage for most users.

ollama pull qwen3.5:27b

This download is roughly 16 GB. It only needs to happen once, and after that the model runs entirely offline.

If your machine has less than 32 GB of RAM, pull a smaller model instead:

ollama pull phi4:14b

For a deeper comparison of every compatible model, see our best local models guide.

Step 3: Install OpenClaw

If you have not installed OpenClaw yet, follow the full installation guide. The short version:

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

Step 4: Point OpenClaw at Your Local Model

Tell OpenClaw to use Ollama as its model provider:

openclaw config set agents.defaults.models.chat ollama/qwen3.5:27b

Verify the configuration:

openclaw models list

You should see your Ollama model listed as the active chat model. No API key prompt will appear because Ollama does not require one.

Step 5: Test Offline

Disconnect from the internet if you want to confirm everything works offline. Then run:

openclaw chat "List all files in my Documents folder"

If OpenClaw returns a file listing, your offline setup is complete.

Which Models Work Offline

Every model you pull through Ollama works offline. The question is which ones handle OpenClaw’s tool-calling protocol reliably.

Avoid models under 7B parameters. They fail tool-calling validation too often to be usable with OpenClaw.

Hardware Requirements

On Apple Silicon Macs, Ollama uses unified memory. A 32 GB Mac can run 27B models without a discrete GPU. The Apple Mac mini M4 is a popular always-on host for this setup.

What Works Offline vs. What Needs Internet

Not every OpenClaw feature works without a connection. Here is the breakdown.

Fully offline (no internet needed):

• File and folder management (create, read, move, rename, delete)
• Code generation and refactoring
• Local system automation and scripting
• Data analysis on local files
• Text summarization and rewriting

Requires internet:

• Web browsing and research skills
• Sending and reading email
• Webhook-based automations (Stripe, GitHub, etc.)
• Pulling new models or updating OpenClaw
• Any skill that contacts an external API or service

The rule is simple: if the task touches only your local filesystem and installed tools, it works offline. If it needs to reach a remote server, it does not.

Cost Comparison

Running OpenClaw with Ollama costs you nothing beyond electricity. For context on what cloud API usage typically costs, see our OpenClaw monthly cost breakdown. Many users start with the free offline setup and add a cloud provider later only for tasks that require frontier-level reasoning.

Troubleshooting

If OpenClaw cannot connect to Ollama, confirm the server is running:

curl http://localhost:11434/api/tags

If that returns a JSON list of models, the server is healthy. If it times out, restart Ollama with ollama serve.

If tool calling fails or the model ignores instructions, switch to a larger model. The 8B models work for simple file tasks but struggle with multi-step workflows. Qwen3.5 27B handles nearly everything reliably.

For more fixes, see the OpenClaw troubleshooting guide.