Introducing NemoClaw: Secure Sandboxing for OpenClaw
Look at your screen. This is OpenClaw running in a locked sandbox. That is what we are going to do in this video with NemoClaw, which is a new product from Nvidia.
At GTC 2026 just today, Nvidia announced something that got the entire OpenClaw community off guard. They built an official security and sandbox layer for OpenClaw agents with the help of OpenClaw's creator, and they open-sourced it on the same day.
Setup Plan: Using vLLM with a Local Qwen Model
I'm going to set it up with vLLM, and then we are going to do it with a local model. Normally, Nvidia is suggesting their own Nemo Tron model, but I'm not going to do that. I'm going to use the OpenShell with this NemoClaw, and we are going to go with this Qwen 3.5 9 billion model, as you can see here.
This is Fahad Mirza, and I welcome you to the channel. If this is the first time you're hearing about OpenClaw and what exactly is this vLLM, don't worry. Just go to my channel, search with OpenClaw. We have done heaps of videos on it. And also, I'm going to use this vLLM tool for it.
If you don't know what vLLM is, how to get it installed, just search for vLLM tutorial and watch this second video, and then you should be able to install vLLM like this, where I already have installed it. And you can see that I am serving this 9 billion model from Qwen 3.5 family on my local system at the moment, and I have shown everything from VRAM consumption to full testing in this video on the same lines.
So, you can see that vLLM is running, model is being served, and now we need to first install NemoClaw. We will also install OpenShell, and then we will see how exactly this thing work.
How the NemoClaw Sandbox Works
Even before that, just allow me a few seconds to walk you through as to what exactly is this diagram and what they have done behind the scene because only then you are going to enjoy this video.
So, first, let me give one-liner on OpenShell. OpenShell is Nvidia's open-source secure runtime for autonomous agents. It runs a hardened sandbox, all locked, built on seccomp, landlock, and network namespaces, where your agent lives. And then, this NemoClaw is the OpenClaw plugin that wires OpenClaw into that sandbox.
Together, they give your agent three guarantees. First, inference is routed through a control provider, the network is locked down by policy, and the file system is isolated. Your agent can only touch what you allow.
When you run OpenClaw, nemoclaw launch, the NemoClaw plugin kicks off a blueprint runner, which I will show you shortly, which is a versioned Python artifact that calls the OpenShell CLI to provision everything in one shot: the gateway, the sandbox container, the inference route, and the network policy. Your OpenClaw agent wakes up inside that sandbox, already configured. So, it means that it cannot write outside, it cannot phone home, it cannot really call your Telegram bank account. Every inference call goes through OpenShell's routing layer to your vLLM backend. This is what is being shown in this diagram. So, hopefully now you understand what exactly is happening here.
System Preparation and Installation
Let me now start installation from scratch so that you could also run it locally.
As far as my system is concerned, I'm using Ubuntu, and this is my GPU card, Nvidia RTX A6000 with 48 GB of VRAM. You can see that due to that vLLM, which is running at the moment, I'm consuming close to 44.3 gig of VRAM for that model.
And if you're looking to rent a GPU just like me on very affordable price, you can find the link to Mass Compute in video's description with a discount coupon code of 50% for a range of GPUs. Make sure to use that coupon. It just makes it so easy. Anyway, let's installation begin. I'm going to first install this OpenShell with this shell command. That is quickly done.
Next up, we need to get clone the repo of NemoClaw. And I will drop the link to this NemoClaw also in the video's description. Once that's done, let's CD into that NemoClaw. Sorry, I'm just... There you go. And now we can run this install.sh script, which is a one-liner script to get everything is installed.
And as mentioned before, make sure that you have Docker installed already. If you don't know how to do that, just search my channel. And it also requires Node. If you don't have the Node, it is going to download and install it. And it takes fair bit of time to do all the checks and stuff, and then it downloads some Docker images.
I think it also runs K3 Kubernetes 3 behind the scene. Don't worry about it. You don't have to learn anything around Kubernetes or Docker, but it takes bit of a time. So, this is what it is doing at the moment in terms of initializing the environment and starting all the services. And everything is almost done. Let's wait. You can just give it any name. I'm just going to press enter here. And it says that it already exists from my previous one. I'll just say recreate it.
Finalizing Installation & vLLM Configuration
And then it is going to take around 5 minutes to create all the containers. And after 5 or 7 minutes, you are going to get this screen. Um just press enter here, otherwise it is going to give you TLS errors.
And our installation is all done. Now, one step which is extremely important is to run these two commands. What these two commands are doing, they are primarily allowing vLLM through firewall from Docker bridge on your local system. So, I already had these rules, so that is why it is skipping, but make sure that you have that thing.
Okay, now we need to set the inference provider to local vLLM with this command. So, I'm just going to run this. Now, you can see that it is configured, otherwise it is going to use the Nemo Tron model. After that, you can simply connect to your sandbox. And you can see that now the prompt has changed. From here, you can test the agent with CLI. I mean, the OpenClaw agent, which is running inside this sandbox. You can ignore these warnings, as you can see that it's just in the alpha stage, and there you go. So, our NemoClaw is running, all set.
Demonstration, Review, and Future Improvements
And it's the same screen which I showed you earlier. There you go. So, it has access to our local system, file system. So, it is telling you what it can do and what it cannot.
And similarly, if you want the familiar OpenClaw interface, all you need to do is to run this command, and you can see that now it has the same, you know, model, you know, TUI, which is running, and then it is going to use your local vLLM model. And still, you know, it might tell you it is using Nemo Tron or model some time, but in effect, it is actually running everything local.
So, that's it. Let me know what do you think about this tool. Pretty good, but I think there are a lot of rough edges, lot of bleeding edges. Taken me around 4 to 5 hours to get it straight, lot of bugs. They also have a bit of a documentation here, which you can check out if you are facing issues. Pretty good, but I think they are still working on it. It was just released today.
I think they could really work on making it a bit more easier. Maybe instead of vLLM, they could just go with Ollama or LM Studio. I think that might be an easier way of doing it, so that people could also use some quantized version. By the way, you can use it with NIM or a few other things, I think. And if you don't know what Nvidia NIM is or Nvidia Nemo 2, by the way, we also have covered NIM and Nemo a lot on the channel in addition to Nemo Tron models. Yep, let me do some shameless promotion here. So, Nemo Tron is here. And what else I mentioned? Uh Nemo, I guess. So, this is our Nemo. Everything is there.
That's it. Let me know what do you think. Please again, like the video and subscribe, and consider becoming a member. Thank you for all the support.
