Agentic AI — NemoClaw on the 8 GB Orin Nano

Your kit can run NVIDIA NemoClaw, an always-on autonomous agent, installed with a single command. This page covers what NemoClaw is, the official install, the agent skills around it, honest 8 GB expectations, and the security decisions it requires.

What NemoClaw is

NVIDIA describes NemoClaw as "a collection of open blueprints for building autonomous agents" — always-on AI systems that reason, plan, and act across real-world workflows. It bundles agent harnesses (OpenClaw, Hermes, LangChain Deep Agents) with NVIDIA Agent Toolkit components: Nemotron models, NeMo, and OpenShell runtime policy controls.

OpenShell is the security layer: "the secure runtime inside it that enforces what the agent can access: files, networks, credentials, and tools."

NemoClaw is alpha software — NVIDIA labels it "Early preview" (since 2026-03-16). Product page: https://www.nvidia.com/en-us/ai/nemoclaw · Build-a-Claw hub: https://www.nvidia.com/en-us/ai/build-a-claw/

Install — the official single command

On the kit, run NVIDIA's installer:

curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash

This installs the default harness, OpenClaw. Two others are selectable with an environment variable:

curl -fsSL https://www.nvidia.com/nemoclaw.sh | NEMOCLAW_AGENT=hermes bash
curl -fsSL https://www.nvidia.com/nemoclaw.sh | NEMOCLAW_AGENT=langchain-deepagents-code bash

On this kit, the installer auto-detects Jetson (Orin and Thor) and applies JetPack host configuration first; on L4T 39.x it loads the br_netfilter module only when missing (without it, the sandbox fails DNS resolution and onboarding hangs at "Setting up OpenClaw inside sandbox"). If you select Ollama, the installer installs it too: "The script also installs ollama (if ollama is selected) so you don't need to manually install it first" (NVIDIA staff). NVIDIA's site documents this device: "Install OpenClaw on Your NVIDIA Jetson Orin Nano" — "a fully local AI personal assistant on Jetson … no cloud APIs needed."

Important — NemoClaw's platform-support matrix (v1.1, 2026-09-04) has no Jetson row; its tested platforms are Linux (Ubuntu 24.04) and DGX OS Spark. Orin Nano support is real in practice — the installer detects the board and NVIDIA documents the flow — but it is not published as formally supported, so expect rough edges.

Requirements that matter here (from NVIDIA's NemoClaw prerequisites page):

Requirement	Min / recommended	On this kit
RAM	8 GB / 16 GB	8 GB total — at the floor
Free disk	20 GB	No built-in storage; use microSD or NVMe (Quick Start)
Node.js / npm	22.19+ / 10+	Install separately
Container runtime	Docker Engine / Desktop / Colima	Socket fix: sudo usermod -aG docker $USER, then newgrp docker

After install — first session

NVIDIA staff point to the Jetson AI Lab walkthrough (vendor guide) for the Orin flow:

1. curl -fsSL https://ollama.com/install.sh | sh — or skip it; the NemoClaw installer can install Ollama too.
2. Pull a 4B-class tool-calling model, such as Nemotron3 Nano 4B (the guide's nemotron-3-nano:30b example targets larger devices).
3. curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
4. Onboard: select Ollama as the model source, and choose the tightest sandbox policy tier that works.
5. source ~/.bashrc, then nemoclaw my-assistant connect; start the agent with openclaw tui.

Agent skills

NVIDIA also ships agent skills — packaged workflows in the open Agent Skills format that extend AI coding assistants (Claude Code, Cursor, Codex) with device-specific automation. Two domains are documented for this era:

• Physical AI (robotics). Isaac ROS ships a catalog of agent skills — per NVIDIA, tasks such as activating the Isaac ROS development container and bringing up the Mission Control cloud stack. The catalog is at https://github.com/nvidia/skills (the "Physical AI" category), installed with npx (Node.js is not part of the standard Isaac ROS environment). Isaac ROS 5.0 adds an isaac-ros-activate CLI and an early-access migrate-node-to-rosidl-buffer skill. See Robotics.
• Video pipelines. The L4T r39.2.1 release notes list "Agent skills for video pipelines" among the What's New items.

One honest gap: the sources for this page document NVIDIA agent skills for Isaac ROS (Physical AI) and for video pipelines; none documents a NemoClaw-specific skill catalog.

Realistic expectations for 8 GB

An always-on agent, a local model, and the Ubuntu desktop do not all fit comfortably on this kit at the same time. The documented budget:

• Usable memory is ~7.6 GB, not 8 GB. NVIDIA: "of the 8 GB physical DRAM, roughly 7.6 GB is usable after firmware and kernel reservations."
• 8 GB is NemoClaw's floor, not a comfort zone. The prerequisites list 8 GB as the minimum and 16 GB as recommended: "On machines with less than 8 GB of RAM, this combined usage can trigger the OOM killer. If you cannot add memory, configure at least 8 GB of swap to work around the issue at the cost of slower performance." This kit's memory is fixed — plan the swap file (Memory Efficiency). The ~2.4 GB sandbox image push has already triggered OOM on an 8 GB Orin Nano.
• The agent and desktop take memory before the model loads. A community guide on NVIDIA's forums puts the OpenClaw runtime at up to ~1 GB; disabling the graphical desktop frees up to ~865 MB (NVIDIA's number), and a community measurement puts GNOME at over 600 MB.
• The oversized-model failure is documented in a community report on NVIDIA's forums. Ollama on an 8 GB board failed to load a 7.4 GB model and a 16 GB model: cudaMalloc failed: out of memory ... failed to allocate buffer for kv cache. File size alone is not the fit test — the KV cache must fit in the same 8 GB.

What fits, per the sources: NVIDIA's validated Ollama defaults (qwen3.6:35b, nemotron-3-nano:30b, qwen3.5:9b) are sized for larger machines; the Jetson AI Lab guide says to start with a 4B-class tool-calling model — "It can work, but do expect weaker performance than the 30B-class models"; and NVIDIA's memory blog puts the tuned 4-bit envelope at LLMs up to ~10B and VLMs up to ~4B parameters — a ceiling for a dedicated setup, not a budget that holds a desktop and agent too.

NVIDIA does not publish tokens-per-second figures for Ollama on this device; treat outside speed claims with care (see Local LLM Inference).

Juxi note: for a workable always-on setup here, plan for headless mode, a 4B-class quantized model, and NVMe storage for the 20 GB requirement and the swap file. That matches what the sources support; anything larger is unverified.

Ollama and agent notes — NVIDIA staff-confirmed

NVIDIA staff debugged the Orin Nano + JetPack 7.2 + Ollama flow on its developer forums, and re-verified Ollama on JetPack 7.2.1 in September 2026.

• Check the GPU first. ollama ps should show 100% GPU in the PROCESSOR column; if it shows CPU, the agent will be very slow.

• Documented failure (June 2026). With NemoClaw + Ollama on a freshly flashed JetPack 7.2 Orin Nano, openclaw tui opened but never answered ("Autocompaction could not recover this turn"). NVIDIA reproduced it: Ollama had skipped GPU discovery (CPU fallback), and the sandbox context window was only 4096 tokens. The staff fix wrote these lines to /etc/systemd/system/ollama.service.d/override.conf:

  Environment="OLLAMA_HOST=127.0.0.1:11434"
  Environment="OLLAMA_CONTEXT_LENGTH=32768"
  Environment="OLLAMA_IGPU_ENABLE=1"
  Environment="GGML_BACKEND_PATH=/usr/local/lib/ollama/cuda_v13/libggml-cuda.so"
  Environment="LD_LIBRARY_PATH=/usr/local/lib/ollama:/usr/local/lib/ollama/cuda_v13"

  then sudo systemctl daemon-reload && sudo systemctl restart ollama; inside the sandbox (nemoclaw my-assistant connect), contextWindow was raised to 32768 in .openclaw/openclaw.json and the config hash refreshed. The reporter confirmed Ollama then ran on the GPU.

• Current status: this workaround should not be needed. Staff, mid-2026: "the issue is fixed in the latest ollama release. The workaround (override.conf) is no longer needed." On JetPack 7.2.1 the upstream installer works and ollama ps reports 100% GPU; the "WARNING: Unsupported JetPack version detected" line is harmless. Test the stock install first.

• If Ollama still falls back to CPU: update Ollama first. A forum user fixed a persistent fallback by deleting the stale /usr/local/lib/ollama/cuda_v12 directory (staff-confirmed removal). Keep override.conf as a fallback of last resort — it is what NVIDIA used successfully on this exact kit.

Security for always-on agents

An always-on agent is a program with credentials and tool access that keeps working while you are not watching. On a device that holds your data, that is a real risk: an agent with tool and shell access here can read, change, or send anything it can reach.

Use the policy layer. NVIDIA describes OpenShell as "the secure runtime inside it that enforces what the agent can access: files, networks, credentials, and tools." During onboarding, pick the tightest sandbox policy tier that still does the job (the Jetson AI Lab walkthrough advises the tightest tier).

Credentials. Give the agent scoped, revocable credentials — dedicated keys and accounts, never your personal ones. Anything the agent can read, it can copy; anything it can use, it can be tricked into using. Messaging integrations act with your identity: NVIDIA's Orin Nano page shows an OpenClaw + WhatsApp example, so use a dedicated account or number.

Network exposure. Keep local services on localhost — NVIDIA's staff configuration for Ollama here binds it to 127.0.0.1 (OLLAMA_HOST=127.0.0.1:11434). Do not expose agent dashboards, control APIs, or model servers to the open internet; for remote access, use a tunnel or VPN you control. The install needs Docker (Engine/Desktop/Colima, per the requirements above) plus a sandboxed container cluster (the OpenShell gateway runs k3s internally) and sudo access.

Operating habits. Start supervised — watch what the agent does before leaving it unattended. Do not give it access you cannot revoke or undo, and keep backups plus a recovery path (see Flashing and Updates). NemoClaw is alpha software ("Early preview"); treat the sandbox as one layer among several, not the only one.

Attention — because this stack runs locally ("no cloud APIs needed"), the security boundary is your device, your network, and your credentials. Review all three before you leave an agent running.

Sources

• NVIDIA NemoClaw product page (checked 2026-09-26) — definition, harnesses, install commands, OpenShell.
• NVIDIA Build-a-Claw resource hub (checked 2026-09-26) — Orin Nano install section.
• NemoClaw — Prerequisites and Platform support (checked 2026-09-26)
• NemoClaw — Troubleshooting (Jetson host configuration) (checked 2026-09-26)
• NVIDIA Developer Forums — NemoClaw on Jetson Orin Super with JetPack 7.2 (NVIDIA staff fix) (checked 2026-09-26)
• NVIDIA Developer Forums — Ollama on Jetson (staff-verified) and JetPack 7.2 GPU acceleration (checked 2026-09-26)
• NVIDIA Technical Blog — Maximizing memory efficiency on NVIDIA Jetson (checked 2026-09-26)
• NVIDIA Developer Forums — AI models that run on Orin Nano Super 8GB (community guide) (checked 2026-09-26)
• Jetson AI Lab — NemoClaw tutorial (vendor guide) (checked 2026-09-26)
• Isaac ROS — Getting Started and Release notes (checked 2026-09-26)
• Jetson Linux r39.2.1 Release Notes (PDF) (checked 2026-09-26) — the "Agent skills for video pipelines" What's New item.

Status: draft, pending review by cheny. Grounded in NVIDIA's official documentation, NVIDIA developer forum posts, and the Jetson AI Lab vendor guide, as of the dates listed; not yet verified on physical hardware by Juxi Technology.

---

NVIDIA® and Jetson™ are trademarks of NVIDIA Corporation. This page is published by Juxi Technology and is not an NVIDIA publication.