NemoClaw + Nebius Token Factory

Run a sandboxed AI agent on NemoClaw (NVIDIA's OpenClaw runtime) backed by Nebius Token Factory inference, with a Telegram interface. No GPU required.

This tutorial walks you through standing up a Price Intel Assistant — a Telegram bot that fetches live product prices from MercadoLibre and returns a structured recommendation, powered by deepseek-ai/DeepSeek-V3.2 on Nebius.

---

Stack

Layer	Technology
Agent sandbox + lifecycle	NemoClaw (NVIDIA, alpha)
Agent runtime	OpenClaw (NVIDIA)
Inference	Nebius Token Factory — OpenAI-compatible
Model	deepseek-ai/DeepSeek-V3.2
Channel	Telegram
Data	MercadoLibre public search API
Host	AWS EC2 m7i.xlarge — Ubuntu 24.04 LTS

---

Prerequisites

• AWS EC2 m7i.xlarge (4 vCPU, 15 GiB RAM), Ubuntu 24.04 LTS
• Nebius Token Factory API key
• Telegram bot token from @BotFather
• Node.js 22.16+ and npm 10+

---

Setup

1 — Install Docker

NemoClaw requires Docker running before its installer starts — it's not bundled.

# Install Docker Engine (Ubuntu 24.04)
# Follow: https://docs.docker.com/engine/install/ubuntu/

# Add your user to the docker group and verify
newgrp docker
docker run --rm hello-world

Hello from Docker! means you're good to go.

---

2 — Install NemoClaw

curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash -s -- --yes-i-accept-third-party-software

Note: the standard curl ... | bash command in NVIDIA's docs fails with a TTY error when piped. Use the flag above. The installer also does not auto-launch the wizard — run it manually after install.

nemoclaw onboard

---

3 — Run the onboarding wizard

The wizard has 8 steps. Key choices:

Inference provider → Option 3 (Other OpenAI-compatible endpoint)

Base URL:  https://api.tokenfactory.nebius.com/v1
API key:   <your Nebius Token Factory key>
Model:     deepseek-ai/DeepSeek-V3.2

Note: Nebius serves all Nemotron models as reasoning-only, which is currently incompatible with NemoClaw's Option 3 path. Use deepseek-ai/DeepSeek-V3.2 or another standard chat model. See issues.md for the full breakdown.

Sandbox name: pick a name, e.g. price-intel

Web search: N

Messaging channels: toggle Telegram, enter your BotFather token

Policy presets: Balanced + enable telegram and github

The first sandbox build takes ~6 minutes.

---

4 — Fix Telegram and verify

Connect to your sandbox:

nemoclaw price-intel connect

Apply this fix before anything else — the NemoClaw wizard writes an invalid Telegram config value that silently crashes the gateway:

sed -i 's/"groupPolicy": "mentions"/"groupPolicy": "allowlist"/' ~/.openclaw/openclaw.json
nohup openclaw gateway > /tmp/gateway.log 2>&1 &

Note: this fix doesn't survive a sandbox rebuild — re-apply it each time. Bug filed with NVIDIA (ISSUE-001).

Test inference:

openclaw agent --agent main -m "hello"

Drop --local if you see it in older docs — that flag is blocked inside NemoClaw sandboxes.

You should get a response. Then DM your Telegram bot — if it replies, the full stack is working.

---

Files

• guide.md — full step-by-step guide with all context and explanations
• issues.md — bugs found during development, filed with NVIDIA and Nebius

---

License

MIT