NVIDIA USDCode MCP Server

// src/server.ts import "dotenv/config"; import { z } from "zod"; import OpenAI from "openai"; const NVIDIA_MODEL = "nvidia/usdcode-llama-3.1-70b-instruct"; const apiKey = process.env.NVIDIA_API_KEY; if (!apiKey) { console.error( "Missing NVIDIA_API_KEY. Set it in your environment or a .env file." ); process.exit(1); } const mcpPath = new URL( "../node_modules/@modelcontextprotocol/sdk/dist/esm/server/mcp.js", import.meta.url ).href; const stdioPath = new URL( "../node_modules/@modelcontextprotocol/sdk/dist/esm/server/stdio.js", import.meta.url ).href; const { McpServer } = (await import(mcpPath)) as any; const { StdioServerTransport } = (await import(stdioPath)) as any; console.error("[usdcode-mcp] Starting server..."); const server = new McpServer({ name: "usdcode-mcp", version: "0.1.0" }); server.tool( "get_usdcode_help", "Ask NVIDIA USDCode for help (Isaac Sim scripting, USD, Python/API tips).\n\nParameters: temperature (0-1, default 0.1), top_p (<=1, default 1), max_tokens (1-2048, default 1024), expert_type (auto|knowledge|code|helperfunction; default auto), stream (boolean; default false). Avoid changing temperature and top_p together.", { question: z.string().describe("Your prompt or question"), temperature: z .number() .optional() .describe("Sampling temperature (0-1). Default: 0.1"), top_p: z .number() .optional() .describe("Top-p nucleus sampling mass (<=1). Default: 1"), max_tokens: z .number() .int() .optional() .describe("Max tokens to generate (1-2048). Default: 1024"), expert_type: z .enum(["auto", "knowledge", "code", "helperfunction"]) // possible values per API .optional() .describe( "Expert to use: auto, knowledge, code, or helperfunction. Default: auto" ), stream: z .boolean() .optional() .describe("Stream partial deltas via SSE. Default: false"), }, async (params: any) => { const { question, temperature = 0.1, top_p = 1, max_tokens = 1024, expert_type = "auto", stream = false, } = params ?? {}; const client = new OpenAI({ baseURL: "https://integrate.api.nvidia.com/v1", apiKey, }); // Build common request payload const request = { model: NVIDIA_MODEL, messages: [{ role: "user", content: question }], temperature, top_p, max_tokens, expert_type, } as any; let text = ""; if (stream) { // Handle streaming by accumulating deltas into a single string const s = await (client.chat.completions.create as any)({ ...request, stream: true, }); for await (const chunk of s as any) { const delta = chunk?.choices?.[0]?.delta?.content ?? ""; if (delta) text += String(delta); } if (!text) text = "No streamed content returned by USDCode."; } else { const completion = await (client.chat.completions.create as any)(request); text = completion.choices?.[0]?.message?.content?.toString() ?? "No content returned by USDCode."; } return { content: [{ type: "text", text }], }; } ); const transport = new StdioServerTransport(); await server.connect(transport); console.error("[usdcode-mcp] Server connected. Waiting for MCP client..."); // Ensure the process stays alive waiting for stdio input process.stdin.resume();