Run OpenClaw on Ollama — free and local.

The trick to running OpenClaw cheaply is recognizing that 60–80% of its requests don't need a flagship model. Heartbeats, status checks, simple lookups, summarization — all of these run great on a local 7B-class model. Reserve cloud calls for the work that actually needs cloud quality.

The pattern is called model routing: Ollama handles the high-volume, low-stakes traffic; Claude Sonnet or GPT-4o handles complex reasoning and the tasks where you'd actually notice quality. We'll wire that up below.

Ollama is a one-line install. It runs as a local server on port 11434 and exposes an OpenAI-compatible API.

# macOS
brew install ollama

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

# Verify
ollama --version
curl http://localhost:11434/api/tags

Use the picker — it'll narrow to the models that actually fit your RAM and use case. Don't just download the biggest one that fits; bigger isn't strictly better, and a model that doesn't tool-call reliably is useless to OpenClaw.

OpenClaw treats Ollama as just another provider — same config shape, different base URL.

{
  "providers": {
    "ollama": {
      "baseUrl": "http://localhost:11434/v1",
      "apiKey": "ollama"
    }
  },
  "models": {
    "default": {
      "provider": "ollama",
      "model": "qwen2.5:7b"
    }
  }
}

Restart the gateway and test:

openclaw gateway restart
openclaw chat
> hello

You should see a response. If not, check openclaw logs --follow while you send a message — the most common issue is Ollama not yet having the model loaded into memory (warm-up takes 5–10s the first time).

The biggest cost win comes from routing traffic by complexity. OpenClaw lets you assign different models to different tasks.

{
  "providers": {
    "ollama": { "baseUrl": "http://localhost:11434/v1", "apiKey": "ollama" },
    "anthropic": { "apiKey": "sk-ant-..." }
  },
  "models": {
    "default": { "provider": "ollama", "model": "qwen2.5:7b" },
    "complex": { "provider": "anthropic", "model": "claude-sonnet-4-6" },
    "heartbeat": { "provider": "ollama", "model": "llama3.2:3b" }
  },
  "routing": {
    "rules": [
      { "match": { "tool": "browser" }, "model": "complex" },
      { "match": { "intent": "long-form" }, "model": "complex" },
      { "match": { "type": "heartbeat" }, "model": "heartbeat" }
    ]
  }
}

• Heartbeats → smallest local model. They run every few minutes and are 80% of token volume. Make them free.
• Browser tasks → cloud. Page snapshots are long and the planning is non-trivial. Worth the cost.
• Long-form drafting → cloud. The quality gap shows up most in extended writing.
• Everything else → local. Most chat, simple tool use, lookups.

We've benchmarked Qwen 2.5 7B against Claude Sonnet on real OpenClaw workloads. The gap is real and predictable.

From actual production setups we've audited:

The bigger your heartbeat overhead, the bigger the savings. Multi-agent setups with frequent check-ins benefit most because that's where the redundant token volume sits.

For the full optimization playbook, read the cost guide.