Ollama 0.33 runs Claude Code on local or Ollama Cloud models
One command points Claude Code — and now Cowork — at open weights on your own GPU or in Ollama's cloud: no Anthropic key, no per-token bill.
Copy markdown
One command, Claude Code on your own model
`ollama launch claude` sets `ANTHROPIC_BASE_URL` to your local Ollama and starts Claude Code on a model you already pulled — `claude --model qwen3.5`. No Anthropic key, no per-token bill, and your code never leaves the machine.
Or borrow a big model from Ollama Cloud
Don't want a 200GB download? `ollama launch claude --model gemma4:cloud --yes` routes Claude Code to an Ollama Cloud model with nothing on disk — a bridge between local privacy and frontier-scale open weights, metered by Ollama instead of Anthropic.
New in 0.33: Cowork and the desktop app
The 0.33 release adds a Claude Desktop hook so Ollama Cloud models can drive Claude Cowork and Claude Code straight from the desktop app, plus in-app Claude model management. Ollama already flags `ollama launch claude-desktop` as transitional and points power users back to the Claude Code setup.
The catch — and a real speed win
Agentic Claude Code needs a model with tool calling, and Ollama recommends 64k+ context for real repos, so a small local model stalls on a big codebase. Upside: the 0.33 line caches resolved model metadata, roughly halving time-to-first-token — 995ms down to 524ms.