Build a NemoClaw agent

NemoClaw is NVIDIA's multi-agent stack, and it is the only option here that is not "one agent, one box." It puts several named agents on one dedicated machine, each inside an NVIDIA OpenShell sandbox with a network egress policy wrapped around it.

It is also, by a distance, the most expensive thing on this site. This guide is mostly about whether you need it, because the honest answer for most people is no.

What you are actually buying

Every agent on this platform already runs in its own Firecracker micro-VM — hardware-level isolation, private network, its own volume. So NemoClaw is not how you get isolation from other tenants. You already have that on a $9 box.

What NemoClaw adds is a second, finer-grained boundary within your own dedicated machine, plus the thing the micro-VM does not give you:

• Egress policy per agent. Each sandbox has a network policy around it, so you can constrain what a given agent is allowed to talk to. On a normal micro-VM the agent can reach the internet; here you decide what "the internet" means for each one.
• Many named agents, one bill. Roughly five sandboxes on the 8 GB box and ten on the 16 GB one, each with its own dashboard and — with distinct bot tokens — its own Telegram front door.
• NVIDIA's stack, specifically. For some teams the answer to "what are you running this on?" needs to be a named vendor's sandboxing and policy layer. That is a legitimate requirement and it is what this plan is for.

The constraint nobody expects

All sandboxes on one box share a single provider and model. This is an NVIDIA limitation, not a configuration step you have missed, and it catches people out.

So a box cannot host a cheap classifier agent alongside a frontier-model research agent. If your reason for wanting several agents on one machine was to mix models per use case, NemoClaw is the wrong shape and several separate micro-VMs are both cheaper and more flexible.

Who should actually pick this

Two profiles, and they are narrow:

1. A team running several agents that must not see each other, where per-agent egress control is a requirement — one per client, one per department, with a real reason they are separated.
2. An organisation whose security review names NVIDIA's sandboxing. If that is on the checklist, this is the plan that satisfies it.

If you are here because "more isolation sounds better," do the arithmetic first. Several agents on separate Plus micro-VMs cost less than one Secure Box, isolate at the hardware level, and let each one run a different model. That is a better deal unless egress policy or the vendor name is the actual requirement.

Building one

1. Work out how many sandboxes you need and size from there. Density is roughly a gigabyte of base plus a few hundred megabytes per idle sandbox, so the sandbox count picks the box.
2. Pick the one provider and model. Every sandbox on the box inherits it, so choose for the most demanding agent — the cheap ones will be fine on a good model, but not the reverse.
3. Name the agents by their job. Sandboxes are named and you will be reading dashboards per sandbox. "support", "research", "ops" beat "agent-1".
4. Give each sandbox its own bot token if you want them separately reachable on Telegram. One token per sandbox is what makes them distinct front doors rather than one shared one.
5. Set the egress policy deliberately. This is the feature you are paying for. An agent that only needs to read your docs should not be able to reach anything else.

Which runtime runs inside

Each sandbox runs OpenClaw by default, or Hermes if the sandbox is running a multi-step workflow rather than a chat agent. So the runtime guides still apply — NemoClaw is the box and the boundary, not the agent itself.

Cost, plainly

The plans are in the fact card and the price is the honest cost of a dedicated machine with that much memory, not a security markup. Note that the free-first-month offer does not apply to the dedicated boxes: a month of a dedicated 8–16 GB machine is a real hardware cost, and we would rather say that than quietly not honour it.

The model bill is separate and, with several agents on one box, adds up faster than people expect. Per-sandbox dashboards exist so you can see which of your five agents is responsible.

Is NemoClaw the right pick?

Not sure? Describe the job rather than the runtime — the builder's solutions engineer reads the same capability matrix this page does and will argue for a different one if it fits better.