Kit: free NVIDIA models in Claude Code (kossolapov.com, 30 Sep 2026)

What is inside:
  nvidia.yaml         the LiteLLM config with four NVIDIA Build models
  nvidia.env.example  the key file template (rename it to nvidia.env)
  claude-nvidia.sh    runs Claude Code on NVIDIA models (macOS, Linux)
  claude-nvidia.ps1   the same for Windows PowerShell
  prompt.txt          a prompt for Claude Code that does the whole setup itself

Fast path: paste prompt.txt into Claude Code while you still have some limit left, and answer its questions.

By hand:
  1. Get an nvapi- key on build.nvidia.com: Login, email, SMS, then Generate API Key on a model page.
  2. uv tool install 'litellm[proxy]'
  3. mkdir -p ~/.litellm and put nvidia.yaml and nvidia.env there (chmod 600 ~/.litellm/nvidia.env)
  4. cp claude-nvidia.sh ~/.local/bin/claude-nvidia && chmod +x ~/.local/bin/claude-nvidia
  5. claude-nvidia      (claude-nvidia -c continues the last conversation)

Models: kimi-k3 (code, the default), deepseek-v4.1-flash (1M-token window), glm-5.3 (agentic tasks), glm-5.3-flash (quick small tasks).
Pick one: NV_MODEL=deepseek-v4.1-flash claude-nvidia

Limits: NVIDIA's trial access is for testing, not production; about 1,000 credits and 40 requests a minute;
NVIDIA may use requests to improve its models, so do not send client code, passwords or personal data.
Anthropic does not officially support running Claude Code on non-Claude models.
