nvidia-claude-proxy

Proxy that exposes Anthropic/Claude-style POST /v1/messages, converts to OpenAI Chat Completions format, and forwards to NVIDIA API.

Visit build.nvidia.com/explore/discover, register an account, and generate an API key. NVIDIA provides hidden models such as moonshotai/kimi-k2.5, z-ai/glm4.7, and minimaxai/minimax-m2.1.

Install

npm install -g nvidia-claude-proxy@latest
# or
npx nvidia-claude-proxy

Initialization

Run the init command to set up your NVIDIA API Key and port interactively. The configuration will be saved to your user home directory (e.g., ~/.nvidia-claude-proxy/config.json).

nvidia-claude-proxy init
# or
npx nvidia-claude-proxy init

Config

The configuration is stored in ~/.nvidia-claude-proxy/config.json (or %USERPROFILE%\.nvidia-claude-proxy\config.json on Windows).

To update your configuration, you can edit this file directly:

{
  "nvidia_url": "https://integrate.api.nvidia.com/v1/chat/completions",
  "nvidia_key": "your-nvidia-api-key",
  "port": 3001,
  "host": "0.0.0.0"
}

Alternatively, you can re-run the init command to overwrite the existing configuration.

| Option | Description | |--------|-------------| | nvidia_url | Upstream API URL (default: https://integrate.api.nvidia.com/v1/chat/completions) | | nvidia_key | Required. Used for upstream authentication. Usually keys starting nvapi- ... | | port | Listen port (default: 3001) | | host | Listen host (default: 0.0.0.0) |

Environment Variables

| Variable | Default | Description | |----------|---------|-------------| | CONFIG_PATH | ~/.nvidia-claude-proxy/config.json | Path to config file | | PROVIDER_API_KEY | from config | Overrides nvidia_key | | UPSTREAM_URL | from config | Overrides nvidia_url | | ADDR | from config / port+host | Full listen address (e.g. :3001, 0.0.0.0:8080); overrides PORT and HOST when set | | PORT | from config | Listen port (used when ADDR is not set) | | HOST | from config | Listen host (used when ADDR is not set) | | UPSTREAM_TIMEOUT_SECONDS | 300 | Upstream request timeout | | LOG_BODY_MAX_CHARS | 0 | Max chars to log in request/response bodies (0 disables) | | LOG_STREAM_TEXT_PREVIEW_CHARS | 0 | Max chars for stream preview in logs (0 disables) | | VERBOSE | false | Set to 1 or true to enable detailed logs |

Run

nvidia-claude-proxy
# or
npx nvidia-claude-proxy

Build & Test

npm run build   # verify modules load (uses config.json.example)
npm test        # run tests (Node built-in test runner)

Usage with Claude CLI

Using moonshotai/kimi-k2.5:

export ANTHROPIC_BASE_URL=http://localhost:3001
export ANTHROPIC_AUTH_TOKEN=any
export ANTHROPIC_DEFAULT_HAIKU_MODEL=moonshotai/kimi-k2.5
export ANTHROPIC_DEFAULT_SONNET_MODEL=moonshotai/kimi-k2.5
export ANTHROPIC_DEFAULT_OPUS_MODEL=moonshotai/kimi-k2.5
claude

Using z-ai/glm4.7:

export ANTHROPIC_BASE_URL=http://localhost:3001
export ANTHROPIC_AUTH_TOKEN=any
export ANTHROPIC_DEFAULT_HAIKU_MODEL=z-ai/glm4.7
export ANTHROPIC_DEFAULT_SONNET_MODEL=z-ai/glm4.7
export ANTHROPIC_DEFAULT_OPUS_MODEL=z-ai/glm4.7
claude

Using minimaxai/minimax-m2.1:

export ANTHROPIC_BASE_URL=http://localhost:3001
export ANTHROPIC_AUTH_TOKEN=any
export ANTHROPIC_DEFAULT_HAIKU_MODEL=minimaxai/minimax-m2.1
export ANTHROPIC_DEFAULT_SONNET_MODEL=minimaxai/minimax-m2.1
export ANTHROPIC_DEFAULT_OPUS_MODEL=minimaxai/minimax-m2.1
claude

API

POST /v1/messages

• Upstream auth: Always sends Authorization: Bearer <nvidia_key> to NVIDIA.

Example (non-stream):

curl -sS http://127.0.0.1:3001/v1/messages \
  -H 'Content-Type: application/json' \
  -d '{
    "model":"z-ai/glm4.7",
    "max_tokens":256,
    "messages":[{"role":"user","content":"hello"}]
  }'

Example (stream):

curl -N http://127.0.0.1:3001/v1/messages \
  -H 'Content-Type: application/json' \
  -d '{
    "model":"z-ai/glm4.7",
    "max_tokens":256,
    "stream":true,
    "messages":[{"role":"user","content":"hello"}]
  }'

POST /v1/chat/completions (OpenAI-compatible)

Same upstream as /v1/messages, but accepts OpenAI request format and returns the response as-is (no conversion). Use this when your client (e.g. Cursor, OpenAI SDK) calls the OpenAI-style endpoint. Upstream auth behaves the same as /v1/messages.

GET /

Health check. Returns {"message":"nvidia-claude-proxy","health":"ok"}.

Notes / Limitations

• Streaming conversion supports delta.content text and delta.tool_calls tool-use blocks.
• Logs may show forwarded request bodies; keep LOG_BODY_MAX_CHARS small and avoid secrets in prompts.
• Requires Node.js >= 18.