Mockup for reviewTech-stack demonstration. Not affiliated with Nebius and not the live Builders Network.About this build →
← Library
WORKSHOP
Official
BEGINNER · 30 min

Running OpenClaw on Nebius

From zero to a live AI agent in 15 minutes. Walks through Token Factory, three deploy methods (npm / Docker / Nebius Serverless), and a hands-on agent build.
openclaw
tokenfactory
Read on the original site ↗

Running OpenClaw on Nebius

30-Minute Workshop Presentation

Event: ClawCamp Workshop Series Duration: 30 minutes (15 min presentation + 15 min hands-on) Level: Beginner Prerequisites: Laptop with browser, Nebius account (free trial), Node.js 18+


SLIDE 1: Title

Running OpenClaw on Nebius From zero to a live AI agent in 15 minutes

ClawCamp Workshop claw.camp/curriculum/agent-inference


SLIDE 2: The Problem

Running AI agents is expensive and complicated

  • A Mac Mini for local inference: $1,000+
  • Exposes your home network
  • Runs ONE agent at a time
  • No scaling, no redundancy, no sleep

What if you could run 100 agents for the cost of one Mac Mini?


SLIDE 3: Liberate Your OpenClaw

You don't need to use Claude or OpenAI.

OpenClaw works with ANY OpenAI-compatible model provider. Today we'll use Nebius Token Factory — open-source models, pay-per-token, no GPU reservation.

Provider Models Pricing
Nebius Token Factory Kimi K2.5, Llama 3.3 70B, DeepSeek-R1, Qwen3, GLM-5 Pay per token
Hugging Face GLM-5, thousands more $2/mo free credits
Local (llama.cpp) Any GGUF model Free (your hardware)

Today's focus: Token Factory. Fast, cheap, production-ready.


SLIDE 4: The Architecture

Orchestration and inference are separate.

Your Agent (OpenClaw)          Token Factory (Nebius)
┌─────────────────────┐       ┌─────────────────────┐
│  Agent logic (CPU)  │──────▶│  Model inference     │
│  Tool calling       │  API  │  (GPU)               │
│  WebSocket gateway  │◀──────│  Llama / GLM / Qwen  │
└─────────────────────┘       └─────────────────────┘
     Runs anywhere               Pay per token
     $0.01/hr                    No GPU needed

Your agent runs on cheap CPU. The expensive GPU work goes through Token Factory's API.


SLIDE 5: Three Ways to Deploy

Method Time Best For
npm install 5 min Local dev, fastest start
Docker 10 min Isolated, repeatable, server deploys
Nebius Serverless 20 min Production, auto-scaling, zero-ops

We'll start with npm (fastest), then show Docker and Serverless.


SLIDE 6: Step 1 — Get Your Token Factory API Key (2 min)

  1. Go to console.nebius.com
  2. Navigate to Token Factory in the sidebar
  3. Click Create API Key
  4. Copy the key

Set it in your terminal:

export NEBIUS_API_KEY="your-key-here"

To persist across sessions, add it to your shell profile:

echo 'export NEBIUS_API_KEY="your-key-here"' >> ~/.zshrc  # macOS
echo 'export NEBIUS_API_KEY="your-key-here"' >> ~/.bashrc  # Linux

This key gives you access to models including Kimi K2.5, Llama 3.3 70B, DeepSeek-R1, and Qwen3. Never commit this key to git.


SLIDE 7: Step 2 — Install OpenClaw (1 min)

npm install -g @openclaw/cli

Verify:

openclaw --version

That's it. No Docker, no containers, no VMs (yet).


SLIDE 8: Step 3 — Connect to Token Factory (2 min)

Option A: Quick onboard (CLI)

openclaw onboard --non-interactive \
  --auth-choice custom-api-key \
  --custom-base-url "https://api.tokenfactory.nebius.com/v1/" \
  --custom-model-id "meta-llama/Llama-3.3-70B-Instruct-fast" \
  --custom-api-key "$NEBIUS_API_KEY" \
  --secret-input-mode plaintext \
  --custom-compatibility openai

Option B: Config file (recommended for production)

Edit ~/.openclaw/openclaw.json:

{
  models: {
    mode: "merge",
    providers: {
      nebius: {
        baseUrl: "https://api.tokenfactory.nebius.com/v1/",
        apiKey: "${NEBIUS_API_KEY}",
        api: "openai-completions",
        models: [
          {
            id: "moonshotai/Kimi-K2.5",
            name: "Kimi K2.5",
            contextWindow: 128000,
            maxTokens: 32000
          },
          {
            id: "meta-llama/Llama-3.3-70B-Instruct-fast",
            name: "Llama 3.3 70B",
            contextWindow: 128000,
            maxTokens: 32000
          },
          {
            id: "deepseek-ai/DeepSeek-R1-0528",
            name: "DeepSeek R1",
            reasoning: true,
            contextWindow: 128000,
            maxTokens: 32000
          }
        ]
      }
    }
  },
  agents: {
    defaults: {
      model: { primary: "nebius/moonshotai/Kimi-K2.5" },
      models: {
        "nebius/moonshotai/Kimi-K2.5": { alias: "kimi" },
        "nebius/meta-llama/Llama-3.3-70B-Instruct-fast": { alias: "llama70b" },
        "nebius/deepseek-ai/DeepSeek-R1-0528": { alias: "deepseek" }
      }
    }
  }
}

Key details:

  • mode: "merge" keeps built-in providers while adding Nebius
  • apiKey: "${NEBIUS_API_KEY}" reads from your environment variable
  • api: "openai-completions" — Token Factory is OpenAI-compatible
  • Model aliases let you switch fast: /model kimi or /model llama70b

SLIDE 9: Step 4 — Start and Verify (1 min)

Start the gateway:

openclaw start

Open your browser: http://localhost:18789

Verify models are loaded:

/models        # List all available models
/model kimi    # Switch to Kimi K2.5
/status        # Check current model and connection

Test the connection directly:

curl -H "Authorization: Bearer $NEBIUS_API_KEY" \
  https://api.tokenfactory.nebius.com/v1/models

Your agent is live. Connected to Kimi K2.5 on Nebius. No GPU on your machine.


SLIDE 10: Model Selection

Pick the right model for your use case:

Model ID Speed Best For
Llama 3.3 70B meta-llama/Llama-3.3-70B-Instruct-fast Fast General agents, tool calling
Kimi K2.5 moonshotai/Kimi-K2.5 Fast Great all-rounder, recommended
DeepSeek-R1 deepseek-ai/DeepSeek-R1-0528 Medium Reasoning, complex chains
GLM-5 zai-org/GLM-5 Very Fast High throughput, simple tasks
Qwen3 Qwen/Qwen3-30B-A3B Fast Code generation, multilingual

Switch models anytime:

openclaw config set agents.defaults.model.primary "moonshotai/Kimi-K2.5"

SLIDE 11: Is Docker Right for Me?

Use Docker when:

  • You want an isolated, throwaway environment
  • You're deploying to a server
  • You need repeatable builds

Skip Docker when:

  • You want the fastest local dev cycle
  • You're just prototyping

Docker Quick Start

# Clone and build
git clone https://github.com/openclaw/openclaw.git
cd openclaw

# Automated setup (handles everything)
export OPENCLAW_IMAGE="ghcr.io/openclaw/openclaw:latest"
./scripts/docker/setup.sh

Or manual:

docker build -t openclaw:local -f Dockerfile .
docker compose up -d openclaw-gateway

Access at http://127.0.0.1:18789/

Health check:

curl -fsS http://127.0.0.1:18789/healthz
curl -fsS http://127.0.0.1:18789/readyz

SLIDE 12: Docker + Token Factory

Connect your Docker container to Nebius. Two options:

Option A: Mount your config file If you already have ~/.openclaw/openclaw.json configured (from Slide 8), Docker picks it up automatically via the volume mount.

Option B: Onboard inside the container

docker compose run --rm --no-deps --entrypoint node \
  openclaw-gateway dist/index.js onboard \
  --auth-choice custom-api-key \
  --custom-base-url "https://api.tokenfactory.nebius.com/v1/" \
  --custom-model-id "meta-llama/Llama-3.3-70B-Instruct-fast" \
  --custom-api-key "$NEBIUS_API_KEY" \
  --custom-compatibility openai

Your agent runs in Docker. Inference runs on Nebius GPUs. Best of both worlds.


SLIDE 13: Structured Tool Calling

The real power: JSON mode for reliable tool calling.

{
  "tool": "get_weather",
  "arguments": {
    "city": "Oakland",
    "units": "fahrenheit"
  }
}

Token Factory supports JSON mode natively — no prompt engineering hacks.

Works with Kimi K2.5, Llama 3.3, and DeepSeek-R1.


SLIDE 14: Going to Production (Nebius Serverless)

When you're ready to deploy for real:

# Install Nebius CLI
curl -sSL https://storage.ai.nebius.cloud/ncp/install.sh | bash

# Deploy OpenClaw on Serverless
nebius msp serverless v1alpha1 endpoint create \
  --name openclaw-agent \
  --image openclaw/gateway:latest \
  --memory 512MB

Auto-scales to zero when idle. Spins up on demand. No VM management.


SLIDE 15: Cost Breakdown

Usage Tokens/day Cost/day Cost/month
Light (100 msgs) ~500K $0.15 $4.50
Medium (1K msgs) ~5M $1.50 $45
Heavy (10K msgs) ~50M $15 $450

Compare: A single H100 GPU = $2-3/hour = $1,500-2,200/month

Token Factory: pay per token, not per hour.


SLIDE 16: Security Checklist

Before going live:

  • Set a gateway token (auth for all connections)
  • Bind to loopback in dev, 0.0.0.0 in production behind a proxy
  • Enable HTTPS via reverse proxy (nginx/Caddy)
  • Set API budget caps in Token Factory
  • Use network security groups to restrict port access
  • For Docker: container runs as non-root user node (uid 1000)

SLIDE 17: Troubleshooting Quick Reference

Problem Fix
Gateway won't start Check port 18789: lsof -i :18789
Token Factory 401 Verify API key: curl -H "Authorization: Bearer $KEY" https://api.tokenfactory.nebius.com/v1/models
"model not allowed" Add model to agents.defaults.models allowlist (see Slide 8)
Model not in /models Verify model exists in BOTH models.providers[].models[] AND the allowlist
Wrong model called Check id matches Token Factory exactly (e.g., moonshotai/Kimi-K2.5)
Docker OOM (exit 137) Increase memory to 2GB minimum
Slow responses Try Kimi K2.5: /model kimi
Connection refused Check bind address and firewall rules

SLIDE 18: What You Built Today

  1. Installed OpenClaw (npm or Docker)
  2. Connected to Nebius Token Factory (open-source models)
  3. Selected a model and configured tool calling
  4. Understand three deployment methods (npm, Docker, Serverless)
  5. Know how to go from prototype to production

Your agent is live. No Claude API key needed. No OpenAI. Just open-source models on Nebius.


SLIDE 19: Resources

Resource Link
OpenClaw Docs docs.openclaw.ai
Docker Install docs.openclaw.ai/install/docker
Token Factory Docs docs.tokenfactory.nebius.com
Token Factory Cookbook github.com/nebius/token-factory-cookbook
ClawCamp Tutorials claw.camp/curriculum
Agent Inference Workshop claw.camp/curriculum/agent-inference
Discord discord.gg/clawcamp

SLIDE 20: Next Steps

Beginner:

  • Deploy via Web Console (20 min) — claw.camp/curriculum/deploy-console
  • Deploy with Install Scripts (15 min) — claw.camp/curriculum/deploy-scripts

Intermediate:

  • Deployment Patterns & Troubleshooting (45 min) — claw.camp/curriculum/deploy-patterns
  • Private Agents for sensitive data (55 min) — claw.camp/curriculum/private-agents

Advanced:

  • Agent Lifecycle & Fine-Tuning (70 min) — claw.camp/curriculum/agent-lifecycle
  • Robotics with Solo CLI (75 min) — claw.camp/curriculum/robotics

SLIDE 21: Thank You

Running OpenClaw on Nebius From zero to a live AI agent in 15 minutes

Open-source models. Pay-per-token. No vendor lock-in.

Questions? Find us at:

Join the personal AI revolution.