OpenClaw + NVIDIA Nemotron 3 Super — Free AI Agent Setup

Run OpenClaw with NVIDIA Nemotron 3 Super for free. The model runs in NVIDIA's cloud — no GPU required on your machine.

What You Get

• OpenClaw — autonomous AI agent that lives on your computer, learns from you, and can run tools
• Nemotron 3 Super — NVIDIA's 120B parameter model (12B active) optimized for agentic tasks
• Cloud inference — model runs on NVIDIA servers, your PC is just the interface
• Web UI + Terminal — use from browser or command line
• Web search — built-in plugin for real-time information
• Free — no API keys, no subscription, $0

Requirements

• Windows 10/11
• Internet connection
• PowerShell 7+
• ~500 MB disk space (for Ollama + OpenClaw, model is in the cloud)

Installation

Step 1: Install Ollama

Open PowerShell and run:

irm https://ollama.com/install.ps1 | iex

Launch Ollama from the Start menu or taskbar. Wait for the tray icon to appear.

Step 2: Pull the Cloud Model

ollama pull nemotron-3-super:cloud

This downloads only a tiny manifest (~345 bytes). The model itself runs on NVIDIA's servers.

Step 3: Launch OpenClaw

ollama launch openclaw --model nemotron-3-super:cloud --yes

This will:

1. Install OpenClaw if not already installed
2. Configure it with Nemotron 3 Super
3. Start the gateway and TUI
4. Open a chat session

Web UI will be available at: http://localhost:18789/#token=ollama

Step 4: Set Thinking Level (Optional)

In the OpenClaw chat, type:

/think high

This enables maximum reasoning depth for complex tasks.

One-Click Launch Script

Save as OpenClaw-Start.ps1 and run with PowerShell 7:

# Maximize window (keeps taskbar accessible)
Add-Type -TypeDefinition @"
using System;
using System.Runtime.InteropServices;
public class Win32 {
    [DllImport("user32.dll")]
    public static extern IntPtr GetForegroundWindow();
    [DllImport("user32.dll")]
    public static extern bool ShowWindow(IntPtr hWnd, int nCmdShow);
}
"@

$hwnd = [Win32]::GetForegroundWindow()
[Win32]::ShowWindow($hwnd, 3) | Out-Null  # 3 = SW_MAXIMIZE

$Host.UI.RawUI.WindowTitle = "OpenClaw + Nemotron 3 Super"
$env:OLLAMA_MODELS = "E:\ollama\models"  # Change to your preferred drive

Write-Host "========================================" -ForegroundColor Cyan
Write-Host "  OpenClaw + NVIDIA Nemotron 3 Super"    -ForegroundColor Cyan
Write-Host "  Cloud Edition"                          -ForegroundColor Cyan
Write-Host "========================================" -ForegroundColor Cyan
Write-Host ""

# Start Ollama if not running
if (-not (Get-Process -Name "ollama" -ErrorAction SilentlyContinue)) {
    Write-Host "[*] Starting Ollama..." -ForegroundColor Yellow
    Start-Process "C:\Users\$env:USERNAME\AppData\Local\Programs\Ollama\ollama app.exe"
    do { Start-Sleep -Seconds 1 } until (
        (Test-NetConnection -ComputerName 127.0.0.1 -Port 11434 -WarningAction SilentlyContinue).TcpTestSucceeded
    )
}
Write-Host "[OK] Ollama" -ForegroundColor Green

# Kill old gateway silently (no flashing windows)
$gwProc = Get-NetTCPConnection -LocalPort 18789 -ErrorAction SilentlyContinue |
          Select-Object -ExpandProperty OwningProcess -Unique
if ($gwProc) { $gwProc | ForEach-Object { Stop-Process -Id $_ -Force -ErrorAction SilentlyContinue } }

# Fresh session (memory lives in files, not chat history)
$sessDir = "$env:USERPROFILE\.openclaw\agents\main\sessions"
if (Test-Path $sessDir) {
    Remove-Item "$sessDir\*.jsonl" -Force -ErrorAction SilentlyContinue
    Remove-Item "$sessDir\sessions.json" -Force -ErrorAction SilentlyContinue
}

Write-Host "[*] Launching OpenClaw..." -ForegroundColor Yellow
Write-Host "    Web UI: " -NoNewline
Write-Host "http://localhost:18789/#token=ollama" -ForegroundColor Blue
Write-Host "========================================" -ForegroundColor Cyan
Write-Host ""

ollama launch openclaw --model nemotron-3-super:cloud --yes

Note: Change OLLAMA_MODELS path if you want models stored on a different drive. The cloud model is tiny, but if you ever pull local models they can be 10-90 GB.

Customizing Your Agent

After first launch, OpenClaw creates config files in its workspace:

File	Purpose
USER.md	Info about you (name, timezone, preferences)
IDENTITY.md	Agent's name, personality, vibe
AGENTS.md	Behavior rules, memory system, tool usage
SOUL.md	Core values and principles
MEMORY.md	Long-term curated memory
memory/YYYY-MM-DD.md	Daily session logs

Edit USER.md and IDENTITY.md to personalize the agent.

Configuration

OpenClaw config: ~\.openclaw\openclaw.json

Key settings:

• Model: agents.defaults.model.primary — which model to use
• Workspace: agents.defaults.workspace — agent's working directory
• Port: gateway.port — web UI port (default: 18789)

Troubleshooting

"Wake up" message appears multiple times

The launch script clears session history before each start. If you removed that, greeting messages accumulate across restarts.

Agent reads too many files on startup

Set agents.defaults.workspace to a dedicated folder, not your entire home directory. The agent reads everything in its workspace on startup.

Plugin warnings about "provenance"

Add the plugin to plugins.allow in config:

"plugins": {
    "allow": ["openclaw-web-search"]
}

Model hangs / slow responses

Normal for cloud models with think high. NVIDIA's free tier may throttle. Try /think low for faster responses.

Ollama installer error "Access denied"

Ollama is already running. Close it from the system tray first, then install.

Architecture

You (browser/terminal)
    ↓
OpenClaw (localhost, your PC)
    ↓
Ollama (proxy on your PC)
    ↓
NVIDIA Cloud (Nemotron 3 Super runs here)

• Your PC = interface only, no heavy computation
• Internet required — no internet = no AI
• PC off = agent off (it's local software connecting to cloud)

Links

• Ollama — model manager
• OpenClaw — AI agent platform
• Nemotron 3 Super on Ollama — model page
• NVIDIA OpenClaw Guide — official setup guide

Credits

Setup guide by @StrangeTeaCreature with help from Claude Code.