Run AI Completely Free and Private with OpenClaw + Ollama

No API key. No monthly cost. Zero data leaves your machine. Here is how to set up OpenClaw with Ollama for a completely local, private AI agent.

Every cloud AI service — OpenAI, Anthropic, Google — processes your requests on their servers. Every message you send your agent, every task you delegate, every document you ask it to analyse travels across the internet to a data centre you do not control. For most personal tasks, this is an acceptable trade-off. For sensitive business data, personal financial information, confidential communications or anything that cannot leave your control, it is not.

Ollama solves this completely. It runs AI models — Llama 3.1, Mistral, Phi-3 and others — directly on your computer. OpenClaw connects to Ollama exactly as it connects to OpenAI or Anthropic, but the model runs locally. Nothing leaves your machine. Not a single character of your messages is sent anywhere.

System Requirements

The hardware you need depends on which model you choose. Llama 3.1 8B (the recommended starting point) runs well on any machine with 8GB of RAM or more. The 70B version of Llama requires 40GB+ of RAM and significantly more storage. Mistral 7B is a good lightweight option for machines with 8GB RAM.

Processing is slower than cloud models — expect responses in 5-30 seconds depending on your hardware and the model. For tasks where privacy matters more than speed, this is a completely reasonable trade-off.

Installing Ollama

Go to ollama.ai and download the installer for your platform, or use the one-liner on Mac and Linux:

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

Then pull your first model:

ollama pull llama3.1      # 8B model — good for most tasks, 8GB+ RAM
ollama pull mistral       # lighter alternative, faster on limited hardware

Configuring OpenClaw to Use Ollama

openclaw config set provider ollama
openclaw config set model llama3.1
openclaw config set telemetry false    # optional: disable usage analytics
openclaw doctor --privacy              # verify: nothing is sent externally

What Works and What Does Not

Local models are excellent for: file management tasks, simple research and summarisation, scheduling and reminders, document processing, personal automation workflows. They are less capable than GPT-4o or Claude on: complex multi-step reasoning, code generation, tasks requiring nuanced understanding of long documents.

For most personal productivity use cases — the ones where privacy matters most — local models perform well. The gap versus cloud models has narrowed significantly with each generation, and Llama 3.1 in particular handles most practical tasks reliably.