NVIDIA Claude Code Proxy

FastAPI gateway that lets Claude Code call NVIDIA's OpenAI-compatible chat API through Claude Code's Anthropic-compatible /v1/messages interface.

Setup

cd D:\test\nvidia-claude-code-proxy
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
Copy-Item .env.example .env
notepad .env

Set at least:

NVIDIA_API_KEY=nvapi-your-key
NVIDIA_MODEL=stepfun-ai/step-3.5-flash

Run

.\start.ps1

Health check:

Invoke-RestMethod http://127.0.0.1:8082/health

Use with Claude Code

In the shell where you start Claude Code:

$env:ANTHROPIC_BASE_URL = "http://127.0.0.1:8082"
$env:ANTHROPIC_AUTH_TOKEN = "local-proxy"
claude

If PROXY_API_KEY is empty in .env, you can omit ANTHROPIC_AUTH_TOKEN. If PROXY_API_KEY is set, ANTHROPIC_AUTH_TOKEN must match it.

Notes

• The proxy always sends requests to NVIDIA_MODEL; Claude Code model names are accepted but ignored upstream.
• /v1/messages/count_tokens uses a local approximation so Claude Code can continue operating.
• Tool calls are translated between Anthropic tool_use/tool_result blocks and OpenAI-compatible tool_calls.
• Streaming responses are converted to Anthropic-style server-sent events.