nvidia-usdcode
A tiny MCP server to access NVIDIA's usdcode model through Claude Code / Codex
claude mcp add --transport stdio ishandotsh-nvidia-usdcode-mcp-server node dist/server.js \ --env NVIDIA_API_KEY="YOUR_KEY"
How to use
This MCP server exposes a single tool backed by NVIDIA USDCode via the NVIDIA Integrate API. The tool, get_usdcode_help, takes a user prompt (question) and returns a single text response from the model. To use it, ensure your server is built and running (npm install / npm ci followed by npm run build, then npm start or node dist/server.js). You can integrate it with MCP clients (e.g., Claude Desktop or OpenAI Codex) by registering the server under your MCP config and calling the get_usdcode_help tool with the required parameters. The primary workflow is: send a question to USDCode, optionally adjust generation parameters (temperature, top_p, max_tokens, expert_type), and receive a helpful USDCode-based reply that can assist with Python snippets, USD scripting, or API usage. The tool is designed for scripting tips and quick API usage examples for Isaac Sim and USD workflows.
How to install
Prerequisites:
- Node.js 18+ installed
- npm (comes with Node.js) and internet access
- NVIDIA_API_KEY from NVIDIA USDCode, available at https://build.nvidia.com/nvidia/usdcode
Step-by-step:
- Install dependencies
ornpm cinpm install - Build the project
This outputs to the dist/ directory.npm run build - Run the server (standalone)
or directlynpm startnode --enable-source-maps dist/server.js - Provide your NVIDIA API key
- Place NVIDIA_API_KEY in a local .env file (loaded automatically) or export it in your shell:
export NVIDIA_API_KEY=YOUR_KEY
- Place NVIDIA_API_KEY in a local .env file (loaded automatically) or export it in your shell:
- Verify the server is accessible via MCP clients using the configured command and arguments (node dist/server.js) and the required environment variable.
Note: The server communicates over MCP via stdio and is intended to be managed by MCP-capable clients. Ensure the environment variable is available to the process at runtime.
Additional notes
Tips and notes:
- If you use a .env file, ensure it is loaded automatically by your environment (e.g., via dotenv or your runtime setup).
- The server exposes a single tool: get_usdcode_help. It accepts parameters: question (string, required), temperature (float, optional, default 0.1), top_p (float, optional, default 1), max_tokens (int, optional, default 1024), and expert_type (string, optional, default "auto").
- In Claude or Codex configs, you will typically provide command node, args [dist/server.js], and env with NVIDIA_API_KEY.
- If you rotate API keys, update NVIDIA_API_KEY accordingly and restart the server.
- Build artifacts (dist/server.js) must be up to date with your codebase for the latest capabilities.
- When deploying in production, consider pinning Node.js to 18.x and locking dependencies with npm ci.
Related MCP Servers
iterm
A Model Context Protocol server that executes commands in the current iTerm session - useful for REPL and CLI assistance
mcp
Octopus Deploy Official MCP Server
furi
CLI & API for MCP management
editor
MCP Server for Phaser Editor
DoorDash
MCP server from JordanDalton/DoorDash-MCP-Server
mcp
MCP сервер для автоматического создания и развертывания приложений в Timeweb Cloud