この本文は AI(Claude)が読むための原文(英語または中国語)です。日本語訳は順次追加中。

NemoClaw

Overview

NemoClaw is an open-source stack by NVIDIA that installs and runs OpenClaw inside a sandboxed environment (OpenShell) with policy-enforced security controls. OpenShell provides Landlock, seccomp, and network namespace isolation. Sandboxes enforce strict egress control — all inference requests route through the OpenShell gateway, not directly to the internet. Network and inference policies are hot-reloadable; filesystem and process policies are locked at creation.

Instructions

1. Install NemoClaw

Prerequisites: Linux Ubuntu 22.04+, Node.js 20+, Docker running, NVIDIA OpenShell installed, NVIDIA API key from build.nvidia.com.

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

The installer runs the guided onboard wizard, creates a sandbox, configures inference (NVIDIA cloud), and applies security policies. After install you see:

──────────────────────────────────────────────────
Sandbox my-assistant (Landlock + seccomp + netns)
Model nvidia/nemotron-3-super-120b-a12b (NVIDIA Cloud API)
──────────────────────────────────────────────────

2. Manage sandboxes from the host

nemoclaw onboard                      # Interactive setup wizard
nemoclaw my-assistant connect         # Shell into sandbox
nemoclaw my-assistant status          # Sandbox health check
nemoclaw my-assistant logs --follow   # Stream logs
nemoclaw start                        # Start auxiliary services
nemoclaw stop                         # Stop services
nemoclaw deploy my-assistant          # Deploy via Brev to remote GPU instance

3. Work inside the sandbox

openclaw tui                          # Interactive chat TUI
openclaw agent --agent main --local -m "hello" --session-id test
openclaw nemoclaw launch              # Bootstrap OpenClaw in sandbox
openclaw nemoclaw status              # Show sandbox health
openclaw nemoclaw logs [-f]           # Stream logs

4. Security policies

Network: All outbound blocked by default, allowlist-based egress, hot-reloadable. When agent requests unlisted host → blocked + surfaced in TUI for operator approval.
Filesystem: Only /sandbox and /tmp writable, locked at creation.
Process: Privilege escalation blocked, seccomp syscall filtering, locked at creation.
Inference: All model API calls intercepted by OpenShell gateway, routed to NVIDIA cloud. Default model: nvidia/nemotron-3-super-120b-a12b.

5. Troubleshoot

nemoclaw my-assistant status                        # NemoClaw health
openshell sandbox list                              # OpenShell sandbox state
nemoclaw my-assistant logs --follow | grep inference # Check inference connectivity

Common issues: Docker not running → start daemon. API key invalid → re-run nemoclaw onboard. Sandbox conflicts → check openshell sandbox list. Network blocked → check egress allowlist.

Examples

Example 1: Set up a new sandboxed coding agent

User request: "I want to run an OpenClaw agent in a secure sandbox with NemoClaw on my Ubuntu server"

Actions taken:

Verify prerequisites: confirm Ubuntu 22.04+, Node.js 20+, Docker running
Install OpenShell from https://github.com/NVIDIA/OpenShell

Run the NemoClaw installer:

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

Follow onboard wizard — enter sandbox name code-agent, select Nemotron model, provide NVIDIA API key
Connect to sandbox:
```
nemoclaw code-agent connect
```
Inside sandbox, start the agent TUI:
```
openclaw tui
```

Expected output: Agent running inside isolated sandbox with Landlock filesystem protection, seccomp syscall filtering, network namespace isolation, and all inference routed through OpenShell gateway.

Example 2: Deploy a sandboxed agent to a remote GPU instance

User request: "Deploy my NemoClaw sandbox to a remote GPU so I can run larger models"