The self-hosted path is the channel's recommended tier once the MaxClaw welcome credit has burned through. It is cheaper than the hosted plan, it gives you 24/7 uptime that no home computer can match, and it sandboxes the agent away from your real filesystem. The three videos in this section cover the flagship one-click install (NemoClaw), the standalone Hermes-on-VPS install, and the Windows WSL2 escape hatch for users who cannot use a Linux host.

If you want your agent to run while you sleep, the VPS path is where it lives.

What you'll learn

  • NemoClaw is a security/privacy wrapper on top of OpenClaw (not a fork), and the install is one line: send nvidia.com/nemo to your existing orchestrator agent and let it do the work.
  • The minimum stack is roughly $3/month for a 4 GB Linux box (Zebra-class) plus a fresh Discord bot token — no credit card on file with AWS required.
  • The four-step Hermes-on-VPS install: pick a plan, grab credentials, SSH in with Termius, run the installer. Total time: under 15 minutes if nothing breaks.
  • Windows is the hardest path: NemoClaw only runs through WSL2 + Ubuntu + systemD, and the GPU passthrough step is where it usually dies. The Nvidia driver must be updated on the Windows side (to 595.79 from March 10, 2025); never install a Linux-side Nvidia driver inside the WSL2 Ubuntu.
  • The fresh-VPS gotchas: git and curl are not always present, and the OpenClaw installation detected. Would you like to import from OpenClaw? prompt has a "yes" path that silently drops cron jobs.
  • Why the channel's recommended provider is Zebra (4 GB / $3/mo), not AWS — every "free VPS" the channel has seen eventually demands a credit card and charges once the trial ends.

NemoClaw Setup Guide: FASTEST Way to Install

This is the second-most-viewed video on the channel for a reason. NemoClaw is an open-source security and privacy layer built on top of the OpenClaw foundation by Peter Steinberger (the OpenClaw creator). It wraps your existing agent with isolated sandboxes, policy-based guard rails, network isolation, data privacy controls, a privacy router, and a hybrid model strategy. The stated goal: solve the fact that "these agents need system access and that creates a lot of risk."

The install is the part people remember. Do not start from terminal. Send your existing OpenClaw orchestrator agent (the channel's is called Stark) the literal line nvidia.com/nemo and let it handle installation. Stark was running on Opus / Sonnet 4.6 for architecture, which is enough to manage the install end-to-end. You do not need to rebuild from scratch — a new Discord bot token plus a new Zebra server is the entire footprint.

The minimum stack:

  • VPS: Zebra at $2/month for 2 GB RAM (basic) or $3/month for 4 GB RAM (recommended for most users). For heavy projects (games, large-scale automation), step up to 8 GB ($5–$7/mo) or 16 GB ($10–$15/mo).
  • Bot token: a fresh Discord bot per NemoClaw agent.
  • Model: the free tier works — nvidia.com/nemotron-3-super on OpenRouter at $0/M input and $0/M output tokens. Alternatives: Hunter Alpha (also free) or minimax-m2.5 / minimax-m2.7 for stronger task execution. Claude Opus / Sonnet 4.6 is recommended for architecture tasks but costs more.

Two important "do nots" from the creator:

  • Do not run on Windows + Docker Desktop. He hit "a lot of problems" attempting a Windows local install and abandoned it. The Windows path is WSL2 — covered in the second video in this section.
  • Do not retrofit NemoClaw onto an existing production agent. "We have a good thing going already." For now, it runs as a separate agent on Discord only, alongside your live stack.

Hermes Agent Setup on VPS

VPS is the cleanest 24/7 home for a Hermes agent — four steps, no local machine footprint. The creator buys a 4 GB / $3 per month plan from Zebra, filtering by price and going standard. For a pure executive-assistant workload, 2 GB at $2/mo is enough. Heavy builds (games, big projects) need 8 GB or 16 GB. He flags that every "free VPS" he has seen eventually demands a credit card and charges once the trial ends, so he treats them as paid plans.

He SSHes in with Termius, not bare terminal. From the Zebra dashboard he grabs the IP address, SSH username (Ubuntu), and SSH password. Termius ships an SFTP pane that shows every file on the VPS. New host → IP → label → Ubuntu → password, and the session is live.

The installer has two fresh-VPS gotchas. If it fails on a brand-new box, run sudo apt update first, because the installer only requires git and curl and many minimal images ship without them. When the prompt says OpenClaw installation detected. Would you like to import from OpenClaw?, he says no for greenfield installs. He warns that the import path he tried elsewhere was "quite troublesome" and dropped the cron jobs, so he now runs a manual migration he says is "99% fixed" — covered in a separate video.

During setup he picks Quick setup over the step-by-step Full mode, registers MiniMax M2.7 with a global-direct API key (Anthropic, Kimi, DeepSeek, Gemini, or OpenRouter work the same way), and defers the messaging-platform step to a follow-up Discord guide. On first launch the agent already exposes tools and skills, including a bundled Polymarket research skill. The VPS route is the "cleanest" because it gives 24/7 runtime, isolates state from local files, and costs a few dollars a month.

NemoClaw WINDOWS Setup Guide (It Actually WORKS)

The Windows path exists, but it is not native. NemoClaw depends on Linux kernel features — landlock, network namespaces — that the NT kernel does not ship. The video's intro states "even experienced developers can't get it to work natively," so the entire guide is built around WSL2 as a Linux sidecar.

The base layer is WSL2 + Ubuntu + systemD. Install WSL2 and Ubuntu in a single PowerShell-as-admin command, reboot, then edit /etc/wsl.conf to add [boot] systemd=true. The creator calls this "the step that almost every guide skips and it'll break everything later." Docker Desktop (Windows AMD64 build from docker.com) is then pointed at Ubuntu 22.04 under Resources → WSL integration and verified with the hello-world image.

The make-or-break step is the Nvidia driver. Before touching Linux, the Windows Nvidia driver must be updated. The rig ships on driver 581.95; the required upgrade is 595.79, released March 10, 2025. Installing the Windows driver auto-writes the GPU stub inside WSL2. The creator is emphatic: "Do not install any Nvidia driver inside the Ubuntu WSL2 side." Doing so overwrites the stub. A repo bug also 404s the dGPU path, requiring a manual swap to Nvidia's generic deb repo.

Then verify GPU passthrough before going further. Run nvidia-smi and nvcc --version. If nvidia-smi returns nothing or errors, the creator says "nothing after this will work" and tells you to stop — not loop. The supported fallback is the agent-install path from the previous NemoClaw video, not a retry.

After install: nemo claw box plan status should show phase ready, then nemo claw box plan connect drops you into the OpenClaw TUI. A hi prompt returning a response confirms the stack is live. Policy presets default to pip and npm only; Discord/Telegram/Slack are deliberately deferred to a future video.

NOTE: not in source video — the nemo claw box plan status and nemo claw box plan connect command syntax is from the NemoClaw Windows setup guide (Course 32 in the archive). The exact command spelling has changed at least once between NemoClaw versions; cross-check with nemo --help on the installed CLI before relying on it.

The four-step VPS install (the clean version)

If you want the cleanest path, do it in this order:

  1. Pick Zebra. 4 GB / $3/mo. Avoid AWS free tier (the channel has heard of bills climbing to roughly $100/month once credits expire). Avoid "free VPS" trials (they convert to paid plans). Monthly billing, cancel any time.
  2. Grab the credentials from the dashboard. IP address, SSH username (ubuntu), SSH password. The SFTP pane in Termius is the difference between a five-minute install and a thirty-minute one.
  3. Run the installer with the fresh-VPS gotchas in mind. sudo apt update first to ensure git and curl are present. Then run the one-line installer. Pick Quick setup over Full. Register a cheap model first (Minimax M2.5/M2.7 or Nemotron 3 Super on OpenRouter). Defer the messaging-platform step.
  4. Test the reinstall OS button before you need it. Click it once on a sandbox server. Confirm the rebuild takes under a minute. You will hit this button in production, and you want to know what it does.

Why VPS is the channel's default tier

Five reasons, all from the channel's coverage:

  1. Cost. A $3/mo VPS is cheaper than MaxClaw's $20/mo plan, cheaper than the electricity for a 24/7 Mac Mini, and cheaper than KiloClaw's $9/mo. The price is "a Starbucks cup of coffee."
  2. 24/7 uptime. Your agent responds even when your Mac Mini is off, the power drops, or your home internet dies. Residential ISPs also block inbound ports, which kills any plan to host agent-built sites from a Mac Mini.
  3. Sandbox isolation. A VPS gives you a clean directory to start from. Share folders in incrementally, not your whole home drive. If the agent wrecks the box, hit reinstall OS and the OS is back in roughly five minutes.
  4. Public ports for agent-built artifacts. The channel's agent Stark runs its own domain and publishes presentations on its own site. That only works because the VPS has public-facing ports. A residential Mac Mini cannot reliably expose services.
  5. One orchestrator agent can provision the rest. Give your main agent the SSH credentials — IP, username, password — and let it set up every other agent. The host told a teammate to "stop buying Mac minis" because the orchestrator can spin up new servers and troubleshoot the workflows on its own.

The bonus: a Minax IO plan layered on the VPS gives roughly 15,000 overnight calls. That's what powers the "I'm going to sleep, build this for me" workflow — the bot runs while you don't.

Try it yourself

  1. Pick Zebra, not AWS. Sign up for a Zebra (or Zeabur / Hostinger) 4 GB / $3-per-month Linux box. Do not give AWS your credit card for an always-on agent workload — the free credits convert to a paid plan the moment they expire.
  2. Delegate the install to the agent. SSH in with Termius, then send the line nvidia.com/nemo to your existing OpenClaw orchestrator agent. Let it provision NemoClaw on the new box. Do not start from terminal.
  3. Patch the fresh-VPS gotchas first. Run sudo apt update so git and curl are present, then answer no to the OpenClaw installation detected. Would you like to import from OpenClaw? prompt. The import path silently drops cron jobs.
  4. Skip Windows + Docker Desktop unless you must. If you must run on Windows, the path is WSL2 + Ubuntu + systemD, with the Nvidia Windows driver updated to 595.79 before you touch Linux. Never install an Nvidia driver inside the Ubuntu WSL2 side.
  5. Verify GPU passthrough before installing the agent stack. On the WSL2 path, run nvidia-smi and nvcc --version. If nvidia-smi returns nothing or errors, stop and use the agent-install path from the NemoClaw video. Do not loop.
  6. Pick a cheap model to start. Default to nvidia.com/nemotron-3-super on OpenRouter ($0/M tokens) for testing. Switch to minimax-m2.5 or Hunter Alpha when you need stronger task execution. Reserve Opus for architecture.
  7. Test the reinstall OS button before you need it. On the Zebra dashboard, click it once on a sandbox server. Confirm the rebuild takes under a minute. You will hit this button in production.
  8. Export SOUL.md, HEARTBEAT.md, and MEMORY.md from any prior MaxClaw install. Drop them in ~/.openclaw/ (or the equivalent Hermes config dir) and the new VPS agent picks up your personality and long-term memory.

Common pitfalls

  • Running NemoClaw on Windows + Docker Desktop. The creator hit "a lot of problems" and abandoned the native path. Use WSL2 + Ubuntu + systemD, and only on Linux hosts with more than 2 GB RAM.
  • Forgetting [boot] systemd=true in /etc/wsl.conf. The creator calls it "the step that almost every guide skips and it'll break everything later." Docker and Kubernetes will silently fail without it.
  • Installing an Nvidia driver inside the WSL2 Ubuntu side. It overwrites the Windows-side stub. Only update the Windows driver (to 595.79 from March 10, 2025) — never the Linux one.
  • Looping on a broken nvidia-smi. If it returns nothing or errors after the toolkit install, stop and switch to the agent-install path from the previous NemoClaw video. The creator says you will be stuck in a loop otherwise.
  • Retrofiting NemoClaw onto a working OpenClaw production agent. "We have a good thing going already." Spin up a separate agent on Discord instead.
  • Picking the "import from OpenClaw" path on a fresh VPS. It silently dropped the creator's cron jobs. Run a manual migration instead.
  • Picking the "import from OpenClaw" path with cron jobs. The import path drops cron jobs by default. The fix is a manual migration (zip the working directory, scp to the new box, reinstall the binary).
  • Using a "free VPS" trial. Every "free VPS" the creator has tried eventually demands a credit card and charges once the trial ends. Treat them as paid plans from day one.
  • Mounting your whole home directory into the VPS. Opus "will do the task at whatever cost, including snooping in every folder." Share folders in incrementally, not the whole drive.
  • Paying for a 16 GB VPS to run an executive-assistant workload. A 2 GB box is enough for a single-purpose agent. Match the RAM to the workload; 4 GB is the recommended default.
  • Skipping the sudo apt update step on a brand-new image. The installer only requires git and curl; many minimal images ship without them. Update first, install second.
  • Picking a desktop app version older than 0.16 for the Hermes front-end. Older builds only show a session token field and will not render the gateway sign-in form. Pin to 0.16+.
  • Relying on the desktop app for 24/7 VPS use. The tmux session that wraps hermes dashboard survives logout roughly 50/50. Use the TUI for terminal or coding work; the desktop app dilutes tokens with UI-schema overhead on every prompt.
  • Burning $30/day on Opus for low-volume work. Minimax's $10/month coding plan runs the same workload for under $1/day. Match the model to the task volume, not the hype.

Sources

  • NemoClaw Setup Guide: FASTEST Way to Install — 31,868 views · video_id: qEFaeLlfLmk — https://youtu.be/qEFaeLlfLmk
  • Hermes Agent Setup on VPS — 924 views · video_id: UbK2kXygPUY — https://youtu.be/UbK2kXygPUY
  • NemoClaw WINDOWS Setup Guide (It Actually WORKS) — 6,773 views · video_id: WBZU-LIduto — https://youtu.be/WBZU-LIduto
  • Supabase query — SELECT video_id, title, views, summary_content, summary_key_takeaways FROM public.videos WHERE video_id = ANY(ARRAY['qEFaeLlfLmk','UbK2kXygPUY','WBZU-LIduto']); against project ttxdssgydwyurwwnjogq.
  • Referenced in coverage: nvidia.com/nemo (NemoClaw install trigger), nvidia.com/nemotron-3-super (OpenRouter free model), minimax-m2.5 / minimax-m2.7 (Minimax executor), Hunter Alpha (free executor), Kimi K2.6 (Kimi's orchestrator), DeepSeek (DeepSeek models), Gemini (Google models), OpenRouter (model aggregator) — verify URLs on the official docs before relying on them.