The MCP hub › How-to › Nemotron 3 Ultra

Give NVIDIA Nemotron 3 Ultra tools for any website

NVIDIA's Nemotron 3 Ultra (launched at Computex, June 2026) is a 500B-parameter MoE model built for long-running agents that reason, keep context, and call tools. But a model only decides which tool to call — you still have to give it real ones. That's what wmcp.sh does: hand it any URL and it returns typed, function-calling-ready tools, so Nemotron can check a price, compare variants, or add to a cart — on sites you never wrote an integration for.

Connect tools to your agent → Grade an MCP server first

~3 lines: Nemotron 3 Ultra + real web tools

# Nemotron 3 Ultra calls real web tools — wmcp.sh supplies them.
import requests
from openai import OpenAI

# 1. Ask wmcp.sh for tools for any URL (Shopify, JSON-LD, or OpenAPI site)
tools = requests.get("https://wmcp.sh/api/v1/tools",
                     params={"url": "https://www.allbirds.com/products/mens-wool-runners"}).json()["tools"]

# 2. Hand them to Nemotron 3 Ultra (NVIDIA's OpenAI-compatible endpoint)
client = OpenAI(base_url="https://integrate.api.nvidia.com/v1", api_key="$NVIDIA_API_KEY")
resp = client.chat.completions.create(
    model="nvidia/nemotron-3-ultra",          # use the exact model id from NVIDIA
    messages=[{"role": "user", "content": "What's the price and is it in stock?"}],
    tools=tools,                                # ← wmcp's function-calling schema
)
# 3. Nemotron picks a tool; resolve the call back through wmcp and feed the result.

wmcp.sh returns OpenAI-/MCP-compatible tool schemas, so it drops into Nemotron's tool-calling loop (or Claude, OpenAI, LangChain, Cursor) the same way. The reads are free; live execution + credential-handling run through the governed proxy.

Connect tools you can trust

An agent that can act is only as safe as the tools you give it. Before you wire a server into Nemotron, check its independent trust grade — wmcp grades every MCP server A–F and watches for rug-pulls.

FAQ

Does NVIDIA Nemotron 3 Ultra support tool calling / function calling?

Yes — Nemotron 3 Ultra is built for long-running agents that reason, keep context, and call tools. But the model only decides which tool to call; you still have to supply the tools. wmcp.sh turns any website or API into agent-callable tools so Nemotron has something real to call.

How do I give Nemotron 3 Ultra access to a website?

Hand the URL to wmcp.sh — it returns a typed list of tools (get_price, add_to_cart, search, etc.) as function-calling / MCP JSON. Pass those to Nemotron's OpenAI-compatible endpoint as the tools array; when it calls one, wmcp resolves it against the live site and returns real data. About three lines of glue.

Is this safe to connect to my agent?

Check the server first: wmcp.sh grades every MCP server A–F for security and watches for rug-pulls. Connect graded, governed tools rather than wiring raw endpoints into a model that can act.