NVIDIA MCP

Local MCP server for NVIDIA NIM and Visual GenAI APIs. It exposes small, practical tools for:

• checking NVIDIA API configuration
• listing available NVIDIA models
• running chat completions through NVIDIA-hosted LLMs
• generating images through NVIDIA hosted Visual GenAI endpoints

This repository is a personal MCP wrapper. It does not contain an NVIDIA API key and does not run image generation locally; the MCP process runs on the local machine and calls NVIDIA-hosted APIs for inference.

Key Loading

The server does not store secrets in this repository.

It loads environment variables from:

1. NVIDIA_MCP_ENV_FILE, when provided
2. this repository's .env, when present
3. process environment variables

Supported key names:

• NVIDIA_API_KEY
• NGC_API_KEY
• NVIDIA_KEY
• nvidia-key

nvidia-key is accepted for compatibility with an existing local .env, but NVIDIA_API_KEY is preferred for new setups.

Codex MCP Config

Example:

[mcp_servers.nvidiaMcp]
command = "node"
args = ["/path/to/minsoo-nvidia-mcp/src/server.mjs"]
env_vars = ["NVIDIA_API_KEY", "NGC_API_KEY", "NVIDIA_KEY"]

[mcp_servers.nvidiaMcp.env]
NVIDIA_MCP_ENV_FILE = "/path/to/.env"
NVIDIA_MCP_OUTPUT_DIR = "/path/to/minsoo-nvidia-mcp/output"

Image Generation Defaults

The current default image endpoint is:

https://ai.api.nvidia.com/v1/genai/black-forest-labs/flux.2-klein-4b

The MCP uses short, direct, one-paragraph prompts best for FLUX-style image generation. Avoid committing generated images unless they are intentionally part of an example or release artifact.

Safety Defaults

Image generation defaults to dry_run=true so a tool call can preview the request without consuming API credits. Set dry_run=false only when you intentionally want to generate an image.

GitHub Safety

Before publishing:

• keep .env untracked
• keep node_modules/ untracked
• keep output/ untracked unless intentionally publishing sample output
• commit .env.example, not a real key file