OpenClaw + Ollama: Run AI Agents for Free with Local Models

What Is Ollama?

Ollama is a lightweight runtime that lets you run large language models on your own hardware. It supports models like Llama 3.2, Mistral, CodeLlama, Phi-3, and dozens more. You download a model once, and it runs entirely on your machine through a local API endpoint.

Think of it as Docker for LLMs. One command pulls a model, another starts serving it. OpenClaw connects to Ollama the same way it connects to Anthropic or OpenAI, but everything stays local and costs nothing to run.

Why Run AI Agents Locally?

There are three strong reasons to run OpenClaw agents on local models instead of cloud APIs.

Step 1: Install Ollama

Ollama is available for macOS, Linux, and Windows. Install it with a single command.

# macOS (using Homebrew)
brew install ollama

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

# Windows
# Download the installer from ollama.com/download

# Start the Ollama service
ollama serve

Ollama runs as a background service on port 11434 by default. Keep it running while your OpenClaw agents are active.

Step 2: Pull a Model

Download one or more models that your agents will use. Each model is downloaded once and cached locally.

# General purpose (recommended starting point)
ollama pull llama3.2

# Fast and efficient
ollama pull mistral

# Optimized for code tasks
ollama pull codellama

# Lightweight, runs on 8GB RAM
ollama pull phi3

# Verify your models
ollama list

Download sizes vary: Phi-3 Mini is about 2.3 GB, Mistral 7B is about 4 GB, and Llama 3.2 8B is about 4.7 GB.

Step 3: Configure OpenClaw to Use Ollama

Point OpenClaw to your local Ollama instance. No API key is needed since everything runs on localhost.

# Initialize OpenClaw (skip if already set up)
npx openclaw init

# Configure Ollama as the model provider
openclaw models auth paste-token --provider ollama

# When prompted for a token, just press Enter (no key needed)

# OpenClaw automatically detects Ollama on localhost:11434

If you changed the Ollama port or are running it on a different machine on your network, update the endpoint in your OpenClaw configuration to point to the correct address.

Step 4: Test the Connection

Create a simple agent and send it a message to verify everything works.

# Create a test agent workspace
mkdir -p agents/test-agent

# Create a minimal SOUL.md
cat > agents/test-agent/SOUL.md << 'EOF'
# Test Agent

## Identity
You are a helpful assistant for testing local model connections.

## Rules
- Keep responses concise
- Confirm which model you are running on when asked
EOF

# Register the agent
openclaw agents add test-agent --workspace ./agents/test-agent --non-interactive

# Send a test message
openclaw agent --agent test-agent --message "Hello! Confirm you are running locally."

# Or start the gateway for web access
openclaw gateway start
# Visit http://localhost:18789

The first message may take a few seconds as Ollama loads the model into memory. Subsequent messages are much faster. On Apple Silicon Macs, expect 20-40 tokens per second with Llama 3.2 8B.

Best Models for OpenClaw Agents

Not all models are equal. Here is what works best for different agent use cases.

Performance Comparison: Local vs Cloud

Local and cloud models have different strengths. Here is how they compare across the factors that matter most.

The Hybrid Approach: Best of Both Worlds

You do not have to choose one or the other. OpenClaw supports per-agent model configuration, which means you can run different agents on different providers in the same team.

The smartest setup is to use local models for routine, high-volume tasks and cloud models for work that demands top-tier output.

# Example hybrid team setup:

# Heartbeat agent (runs every 5 min, checks system health)
# → Use Ollama/Llama 3.2 (high volume, simple task, $0 cost)

# Research agent (summarizes articles, extracts data)
# → Use Ollama/Mistral (routine processing, no API cost)

# Content writer (creates blog posts, marketing copy)
# → Use Claude Sonnet (complex creative work, quality matters)

# Code reviewer (analyzes PRs, suggests improvements)
# → Use Ollama/CodeLlama (code-specific, runs locally)

Each agent's SOUL.md can specify which model provider and model to use independently. This means your heartbeat agent running 288 times per day costs nothing, while your content writer uses Claude only when it has actual work to do.

Hardware Requirements

What you need depends on which model you plan to run. Here are the practical minimums.

Frequently Asked Questions