NemoClaw VPS Setup — Install, Onboard and Connect Safely (2026)

What you're building

• OpenClaw — the agent you talk to. NemoClaw also supports Hermes and LangChain Deep Agents Code; this guide uses OpenClaw, the default.
• OpenShell — NVIDIA's sandbox. The agent runs inside it; network egress goes through a policy you control, and provider credentials stay on the host — the sandbox gets a placeholder that OpenShell swaps for the real key at egress.
• NemoClaw — the CLI that installs, onboards and operates both, with commands such as nemoclaw <sandbox> status.

Known issues right now (checked 2026-09-25)

OpenClaw 2026.9.1+ sandboxes fail with a legacy credential-migration error

OpenClaw 2026.9.1 rejects the legacy auth profile NemoClaw generates for managed inference.local routes, and the repair OpenClaw suggests can't run while the Gateway holds its maintenance window, so agent turns fail on fresh managed sandboxes. The fix (PR #12237) clears the host NVIDIA credential aliases early for managed routes and removes only NemoClaw's own generated legacy profile. It merged on 2026-09-24, after v0.0.129, so it ships in the next release. Until then: stay on your current install if it works, or update NemoClaw once the next version is out. Direct (non-managed) routes are unaffected.

Dashboard port refused in the 8642–8652 range

NemoClaw reserves ports 8642 through 8652 for each Hermes sandbox's OpenAI-compatible API, so it rejects any port in that range as a dashboard port — for every agent, not just Hermes. Pick a port outside it.

Step 1 — A server that meets the minimum

Any Linux VPS works; the tested server path is Linux with Docker. Check the plan against NVIDIA's requirements:

Check the vCPU count — entry-level VPS plans often have 2. With less than 8 GB of RAM, pushing the ~2.4 GB sandbox image can trigger the out-of-memory killer; NVIDIA's workaround is at least 8 GB of swap. Native Windows isn't supported; use Linux, macOS or WSL.

Firewall: allow inbound SSH only. Keep the dashboard and OpenShell gateway ports closed — you'll reach them through SSH. The host must still allow outbound DNS, HTTPS, container registries, package registries and your inference provider. The OpenShell policy governs traffic from inside the sandbox; it doesn't replace the host firewall.

Step 2 — An inference credential

• Free: an NVIDIA API key from build.nvidia.com (profile → API keys).
• Or an Anthropic, OpenAI, OpenRouter or Gemini key if you'd rather run Claude, GPT or another model. You pick the provider during onboarding.

The installer asks for the key in its own prompt. Don't paste keys into an agent chat, and don't put them on a command line where they land in shell history.

Step 3 — Install and onboard

SSH into the server and run the hosted installer:

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

If Docker is missing, the installer installs it (via Docker's official convenience script, with a sudo prompt) and adds your user to the docker group. It needs Node.js 22.19+ and npm 10+ for the NemoClaw CLI and can install Node itself. Then the wizard:

1. Accept the third-party software notice.
2. Choose OpenClaw as the agent.
3. Choose an inference provider and model, and enter its credential when prompted.
4. Press Enter to accept the suggested sandbox name, my-assistant (the commands below assume it).
5. On a first run, skip optional web search and messaging, and accept the suggested network policy tier.

If onboarding stops partway — a long run over SSH can drop — fix what it reports and resume rather than starting again:

nemoclaw onboard --resume

Step 4 — Confirm it's actually ready

Don't treat "the process is running" as ready. Check:

openshell sandbox list                      # my-assistant should be Ready or Running (NotReady is not ready)
nemoclaw my-assistant status                # inference route should report healthy
nemoclaw my-assistant connect --probe-only  # should exit 0

Step 5 — Open the dashboard through an SSH tunnel

On the server, print the tokenised dashboard URL:

nemoclaw my-assistant dashboard-url --quiet

On your own computer, open a tunnel to the server's loopback port, then open the printed URL in your browser:

ssh -N -L 18789:127.0.0.1:18789 <user>@<server-ip>

18789 is the default dashboard port, but NemoClaw may pick the next free one up to 18799 — use the port in the printed URL, on both sides of the -L. Don't open port 8080 for remote access, and don't bind the dashboard to every interface when a tunnel does the job. The URL contains a token: keep it out of logs, shell history and screenshots.

Prefer the terminal? These run on the server:

nemoclaw launch my-assistant   # preflight, then the OpenClaw TUI
nemoclaw my-assistant connect  # a shell inside the sandbox

Step 6 (optional) — Switch to Claude, GPT or another model

If you onboarded with one provider and want another, the provider must be registered first. List what's registered, then switch:

nemoclaw credentials list
nemoclaw inference set --provider anthropic-prod --model claude-opus-5-5 --sandbox my-assistant

Canonical provider IDs include nvidia-prod, anthropic-prod, openai-api, openrouter-api, gemini-api, ollama-local and vllm-local. If the provider you want isn't registered, run nemoclaw onboard to add it, then switch. The switch restarts only the OpenClaw gateway (a brief interruption), and NemoClaw validates the new route from inside the sandbox before it commits. Model IDs change often — check the cost calculator for current ones.

Step 7 (optional) — Connect Telegram

nemoclaw my-assistant channels add telegram
nemoclaw my-assistant channels status

channels add prompts for the bot token (create the bot with BotFather first), registers it with the OpenShell gateway rather than the sandbox, and offers to rebuild so the image picks up the channel. Discord, Slack, WhatsApp, WeChat and Microsoft Teams are also available; WhatsApp, WeChat and Teams are experimental, and Teams needs a public HTTPS endpoint ending in /api/messages — the one case where you do need to expose something.

After a reboot

NemoClaw doesn't yet start everything automatically after a host reboot. NVIDIA's recovery order:

sudo systemctl start docker && docker info
nemoclaw my-assistant status      # reports which layer is down
nemoclaw my-assistant start       # if the sandbox container is stopped
nemoclaw my-assistant connect --probe-only
nemoclaw my-assistant recover     # if the agent gateway or host forward is still unhealthy

Then restart your SSH tunnel. Stopping, starting and reconnecting a sandbox keeps its home and workspace in place, so a reboot doesn't need a restore step. To replace the software, use nemoclaw my-assistant rebuild, which preserves the complete native home and workspace. nemoclaw my-assistant destroy removes the sandbox's managed state, so download anything you need first. nemoclaw backup-all is the installer's safety backup before an upgrade, written to ~/.nemoclaw/rebuild-backups/; NVIDIA states it is not a general snapshot create, list or restore interface. Details are in NVIDIA's Persist Sandbox State guide (an earlier version of this page told you to run snapshot create; the current command reference no longer lists it).

What's next

• Tighten the network policy — NemoClaw denies outbound traffic by default; add only what the agent needs. nemoclaw my-assistant policy add --from-file <preset>.yaml --dry-run previews a change. See our NemoClaw policy guide.
• Skills — OpenClaw skills database and writing your own.
• Hardening — OpenClaw security applies inside the sandbox too.
• Updates — check with nemoclaw upgrade-sandboxes --check; don't run openshell self-update on a NemoClaw-managed host — re-run nemoclaw onboard instead.
• Safe teardown — nemoclaw uninstall keeps your sandbox backups and registry (rebuild-backups/, backups/, sandboxes.json) by default; add --destroy-user-data only for a full purge.