NemoClaw
NemoClaw is NVIDIA's open-source way to run an always-on agent (OpenClaw, Hermes Agent or LangChain Deep Agents Code) inside an OpenShell sandbox. The sandbox can reach only the network endpoints you allow, and your model-provider keys stay outside it on the OpenShell gateway. NemoClaw handles install, onboarding, channels, rebuilds and backups. You don't need an NVIDIA GPU: most people run it on a Linux VPS with a hosted model. NVIDIA still calls it early-preview software.
Three agents: OpenClaw, Hermes Agent and LangChain Deep Agents Code, each with its own sandbox image and baseline policy. Policy tiers at onboarding (Restricted, Balanced by default, Open, Personal) plus reviewed presets you add with nemoclaw <sandbox> policy add. Managed MCP servers with credentials kept outside the sandbox, backup-all before upgrades, and recover and rebuild for broken sandboxes, and a Windows WSL2 / rootless Podman path. OpenClaw 2026.9.1 users on managed routes should read our known issue. All six guides below were re-checked against NVIDIA's docs on 2026-09-27.
Guides
policy add presets (GitHub, Gmail, Slack…), approving blocked requests in openshell term, and custom presets.nemoclaw inference set. Provider IDs, compatible endpoints, the one-route-per-gateway rule, and verification.nemoclaw <sandbox> skill install, ClawHub skills, plugins vs skills, and the presets a skill needs to reach the network.Skills inside a NemoClaw sandbox are ordinary OpenClaw (or Hermes) skills. Writing them is covered in the platform guides:
→ Skills Guide: Write Your Own
→ Skills: Finding & Vetting on ClawHub
Install a local skill from the host with nemoclaw <sandbox> skill install ./my-skill/. A skill that calls an outside service also needs its policy preset. See Skills on NemoClaw.
NemoClaw vs OpenClaw — What's Different
| Feature | NemoClaw | OpenClaw |
|---|---|---|
| Execution environment | OpenShell security sandbox (containerised) | Direct on host OS |
| Network policy | Deny-by-default; you allowlist domains explicitly | Permissive by default |
| API key security | Keys stored outside sandbox via OpenShell provider system | In config files on host |
| Agents | OpenClaw, Hermes Agent or Deep Agents Code | OpenClaw |
| Setup | One-line installer plus onboarding wizard (installs Docker if needed); dashboard reached over an SSH tunnel | Single CLI install |
| Skill compatibility | Same skills; network-using skills need a policy preset | Full |
If you don't need the sandbox security model, OpenClaw is simpler to start with. If you're handling credentials, production infrastructure, or sensitive data, NemoClaw's defaults are worth the extra setup.
NemoClaw Use Cases
NemoClaw shines for self-hosted, containerised setups where you control the GPU and the policy.
- Daily journal — privacy-first, can run entirely on local Ollama
- Code review automation — bound its reach with policy presets
- Dependency updater — runs in your container, no cloud egress
- All 12 use cases →
NemoClaw Troubleshooting
- Provider API key invalid — per-profile credential reset
- All troubleshooting entries →
NemoClaw Security
- Skill allowlisting — OpenShell network policy is the enforcement layer
- Sandboxing — containerisation and read-only volumes
- Secrets & credentials — local model = secrets never leave your network
- 15-minute hardening checklist
Related on This Site
- OpenClaw hub — the base framework NemoClaw runs on top of
- IronClaw — a different security-hardened approach: stricter defaults without Docker/OpenShell
- Decision guide — pick the right agent for your use case
- Weekly News Digest — NemoClaw updates and OpenShell security advisories
Official sources
Checked against the NVIDIA NemoClaw documentation and command reference (v0.0.129) on 2026-09-27. If anything here disagrees with these pages, trust the vendor docs.