Install NVIDIA NemoClaw v0.0.69: OpenClaw Deploy Guide

Here’s the unpopular opinion: NVIDIA’s “one command install” marketing is technically true and practically misleading. The curl-pipe-bash finishes in minutes. The interesting failures start after – port conflicts on 8642, a Hermes policy that quietly stops trusting GitHub, an uninstaller that leaves half your host untouched. If you’re deploying NVIDIA NemoClaw to run OpenClaw agents in a sandbox, plan for the friction, not the demo.

This guide covers NemoClaw v0.0.69 (current release as of mid-2026) and pins down the exact spots where the docs go quiet.

What NemoClaw actually is – and why version discipline matters

NemoClaw is an open source reference stack for running always-on AI agents inside NVIDIA OpenShell sandboxes. Single CLI. Guided onboarding, hardened blueprint, routed inference, network policy, lifecycle management – all of it. Think of it as the host-side installer and orchestrator; OpenClaw is the agent itself. (NVIDIA/NemoClaw on GitHub)

Announced by Jensen Huang at GTC on March 16, 2026, the project is available in early preview – and the docs say it plainly: not production-ready, interfaces may change without notice. (TechCrunch GTC 2026 coverage) Pin your version. Expect breakage on upgrade. Don’t put this in front of paying customers yet.

System requirements

Officially, NemoClaw targets NVIDIA RTX Spark laptops, GeForce RTX PCs and laptops, RTX PRO workstations, and DGX Station or DGX Spark hardware. macOS and Linux hosts work. Windows support via WSL is not officially documented in current NemoClaw materials – check the repo README for the latest platform guidance before assuming it works.

One thing first: get your NVIDIA API key

Grab a key from build.nvidia.com before running anything. NVIDIA serves open models free through an OpenAI-compatible API at https://integrate.api.nvidia.com/v1, authenticated via that key. The default model in OpenClaw’s NVIDIA provider is Nemotron 3 Ultra – 550B total parameters, 55B active, with a 1M-token context specification. (Per OpenClaw NVIDIA provider docs, as of v0.0.69.)

Export it as a variable. Don’t use the flag:

export NVIDIA_API_KEY="nvapi-..."

The --nvidia-api-key flag lands the key in shell history and ps output. A key in ~/.bash_history is a key on your attacker’s laptop. The environment variable costs nothing and leaks nothing.

Install NemoClaw v0.0.69

The official installer lives at the NVIDIA/NemoClaw GitHub repo. Before you run it, pick your agent variant:

# Hermes variant - must be set BEFORE the installer runs
export NEMOCLAW_AGENT=hermes

# Run the official installer from the repo (check current command in README)
# then:

nemoclaw onboard # prompts for NVIDIA API key
nemoclaw launch my-assistant

NEMOCLAW_AGENT is not optional if you want Hermes specifically – set it before install, or use the nemohermes alias afterward. For LangChain Deep Agents or plain OpenClaw, the default applies.

Onboarding writes your key to ~/.nemoclaw/credentials.json (per community installer documentation quoting NVIDIA’s developer guide). Lock the permissions immediately: chmod 600 ~/.nemoclaw/credentials.json.

Verify the install – both layers

Most tutorials stop at “launch succeeded.” That’s not enough. Errors can come from either NemoClaw or the OpenShell layer underneath – and the status checks are different commands for each:

# NemoClaw layer
nemoclaw my-assistant status

# OpenShell layer - the sandbox itself
openshell sandbox list

# Connect and test
nemoclaw my-assistant connect

Green nemoclaw status but nothing in openshell sandbox list? The problem is at the runtime layer, not the orchestrator. Start there: nemoclaw my-assistant logs --follow.

Successful launch prints a banner with sandbox name, security profile, and model – something like Sandbox my-assistant (Landlock + seccomp + netns) - Model nvidia/nemotron-3-ultra (NVIDIA Endpoint API).

The three failures that actually show up

These aren’t theoretical. They appear repeatedly in installs and aren’t obvious from the output.

1. Homebrew or Git operations hang or 403 from inside the sandbox

As of v0.0.69, GitHub was removed from the Hermes baseline policy allowlist. The NemoClaw release notes are explicit: Homebrew Git operations require the GitHub policy preset. Add it:

nemoclaw policy-add github --sandbox my-assistant

(Exact command syntax – verify against current docs, as policy CLI may evolve in alpha.)

2. First response never arrives – the ~120s hang

Large custom NVIDIA-hosted models can take longer than the default ~120s model idle watchdog before emitting a first response chunk. It’s not a crash. The fix: raise the idle timeout in your config on the specific model row, not globally. A global increase masks real hangs elsewhere.

3. Port 8642 already bound

NemoClaw reserves Hermes port 8642 across all agent variants – by design, not accident. There is no runtime override. If something already binds 8642, the sandbox gateway fails to start. Check before you install:

lsof -i :8642

If something answers, move it first. This is the first thing to run before any nemoclaw launch.

Alpha discipline: Snapshot your config to git before every NemoClaw upgrade. Schema changes happen without notice. If onboard fails after an upgrade, your git diff is the fastest way to see what changed – faster than re-reading release notes from scratch.

A brief pause on what the sandbox model actually means

There’s a small idea inside NemoClaw worth sitting with. The stack treats every agent action as a policy decision made at runtime – not a permission granted once at install. That’s why the release notes read like a firewall changelog, and why removing GitHub from the default allowlist is a feature, not an oversight. Whether that level of control is overkill for a hobby project or the minimum bar for anything touching real data is a question this alpha stage doesn’t answer yet. Worth thinking about before you open up network policies wholesale.

Upgrading and uninstalling

Upgrades on alpha software: a stability gamble. The blueprint lifecycle follows four stages – resolve the artifact, verify its digest, plan the resources, apply through the OpenShell CLI. A bad artifact should fail at verify rather than corrupt a running sandbox. Still, tear down sandboxes before upgrading the CLI. The alpha warning from the top of this guide applies double here.

Uninstall path (check the current command in the repo README – exact URLs shift in alpha):

# Get the current uninstall script from github.com/NVIDIA/NemoClaw
curl -fsSL [current-uninstall-url] | bash

The script removes sandboxes, the NemoClaw gateway and providers, related Docker images and containers, local state directories, and the global nemoclaw npm package. What it does NOT remove: Docker, Node.js, npm, Ollama. If you installed those specifically for NemoClaw, that cleanup is manual. Plan for it before you run the script expecting a clean slate.

FAQ

Do I need a DGX Spark to run NemoClaw?

No. The NVIDIA cloud endpoint handles inference via API key – no local GPU required.

What’s the difference between OpenClaw and NemoClaw?

OpenClaw is the agent framework – you could run it bare on a laptop for a quick chat session. NemoClaw adds the installer, OpenShell sandbox, network policy layer, and inference router around it. The moment your agent needs to touch Slack, a filesystem, and a shell simultaneously, NemoClaw’s wrapping is what keeps that contained. Without it, you’re managing all of that yourself.

Is my API key safe in ~/.nemoclaw/credentials.json?

Stored as plaintext – NemoClaw does not encrypt the file at rest. chmod 600 ~/.nemoclaw/credentials.json is the immediate fix. Don’t let your home directory sync to a shared backup. If you suspect exposure, rotate the key at build.nvidia.com – takes 30 seconds and invalidates the old key immediately. The file permission is the thing most people skip and regret.

Before you install anything: run lsof -i :8642. Generate your API key at build.nvidia.com. Read the current NemoClaw release notes – policy defaults shift often enough that last month’s tutorial may already be wrong.