Claude Cowork + Ollama: 100% Free and Private Local Setup

Key Takeaways

• Claude Cowork's API base URL is configurable — change it from api.anthropic.com to http://localhost:11434/v1 and it routes all requests through Ollama.
• Recommended models for this setup: Qwen 3.6, Gemma 4, or Llama 3.3. Larger models (30B+) give significantly better skill execution quality if your hardware supports them.
• Local models lack Anthropic's Claude system prompt tuning, so some Cowork skills may need minor prompt adjustments to work reliably.
• Ollama must be running before launching Cowork — start it with ollama serve in a separate terminal.
• For privacy-sensitive workflows (legal documents, personal finance, confidential client data) this setup is the only way to use Cowork without sending data to a cloud provider.
• Performance bottleneck is your GPU. An RTX 4090 handles 30B-parameter models at ~40 tokens/sec — fast enough for practical daily use.

Commands & Code Mentioned

ollama pull qwen3:6b
ollama pull gemma4:latest
ollama serve
# In Cowork settings, set API base URL to:
# http://localhost:11434/v1

When to Use Local vs Cloud Models with Cowork

Local Ollama models are best when you have sensitive data, want zero ongoing cost, or need to work offline. They work well for structured, well-defined tasks: summarising documents, formatting data, generating draft content. Where they fall short is on complex multi-step reasoning, long context windows, and tasks that require up-to-date knowledge.

Cloud models (Claude Sonnet, Claude Haiku via Anthropic) give much better results on open-ended agent tasks, code generation, and anything requiring tool use across many steps. A pragmatic approach: local models for your routine daily Cowork tasks, cloud API for your most demanding agent workflows.

Related on OpenClawDatabase

• Claude Cowork Setup Guide — official setup walkthrough including API key configuration
• Claude Cowork Pricing — compare local vs cloud costs at different usage levels
• Ollama + MCP: Free Private Local Models with Full Tool Use — related local model setup for OpenClaw