NemoClaw захищає ваші AI‑агенти: безпека без компромісів

Nemo claw the must-h have solution for your open claw setup. Enterprisegrade security. 300 plus malicious skills identified in open claw hiding inside AI marketplace resulting in API key stolen browser credential exposed. That's why Nemo claw is introduced to secure your open claw. Open claw is inside Nemo claw and when a thread comes it automatically gets blocked. When there's a mass attack automatically everything is getting blocked. When the approved request come automatically it goes through. All the log is getting recorded. Nemo claw is denied by default. That means anything which is going out and coming in is blocked. So you need to know how to configure Nemo claw so that when a request go out instead of getting blocked you want to apply a policy and then allow the request to go out of Nemo claw. This is what we are going to try to achieve. Looking at what went wrong before, what Nemo claw actually is, the three key components of Nemo claw, API key security model, how to set it up locally in your DJX Spark machine, adding your network policies, realtime monitor with OpenShell dashboard, as you can see here, going through the logs, and many other things. That's exactly what we're going to see today. Let's get started. >> [music] >> Here is my nano claw running. Previously when I asked what is the weather in London, it was not able to answer that because as you can see openshell was blocking it. So then I allowed through openshell by changing few configuration as a result again it was able to fetch the weather for me. So I'll take you through how you can set up nemo locally and then allow these requests. How this is by default very secure. Previously what you do in regards to open claw you install a skill. Skill runs silently and the data leaves quietly. The fake skill automatically sends all your API credentials and your passwords to remote machine. But with Nemo claw this is prevented. So what is Nemo claw? Is that a new agent? No. It runs open claw inside. Same agent, same brain, same skills, just secured. Does it compete with open claw? It wraps open claw. Is it locked to Nvidia Nimatron? No, it can use all other models. Does it require Nvidia GPU? It is full hardware agnostic. So no, it doesn't need a NIDA GPU. Open claw runs in a sandbox protected by guardrails with access control, privacy policy and only approved skills can run. It is a private inference router to route between local models and the frontier models more securely. Think it like a space mission where open claw is the astronaut and the open shell is a spacecraft and the nemo claw is the mission control which builds the mission, sets flight rules, monitors every tele signal, logs all decisions. Nothing launches without approval. All the request are tracked, monitored and then blocked if it's a malicious code. If it is safe then it automatically allow the request and perform the task. So if you don't know about open claw it basically consist of five part engine. One is gateway that is the one which runs 24/7 with the user interface. Behind the scenes you got the reasoning engine that is the large language model. Then you got the memory for the agent. So the large language model is nothing but agent. Memory for the agent skills for the agent and schedule task for the agent. And these all connects through WhatsApp, Telegram, Slack and much more. So you send a message through any of these platform and it automatically respond to your request. So the key advantage is how the API keys stay safe. So agent is running inside the sandbox. So each API request goes through internal host that is inference.local. Then it goes to open shell gateway which automatically blocks any personal data and then it's passed to AI provider such as open AI anthropic gemini and grock keys never enter the sandbox automatic key rejection and breach proof by design. So what's the difference between Nemo claw and plain open claw? Plain open claw full user account permission. It can read everything but with Nemo claw it's restricted to two folders sandbox folder and the temporary folder. Open claw network calls no restriction calls anywhere but with Nemo claw you can whitelist and blocks at OS level by default. API key handling for open claw it sends directly with raw request but in Nemo claw keys are intercepted by gateway never enter the sandbox. Privacy filtering in open claw raw requests send assis but with Nemo claw personal data stripped before requests leave the sandbox audit logging no logging or alerting but with Nemo claw every allow deny decision logged real time user interface so now we are going to set up Nemo claw inside my DGX spark machine which is running locally within my own network and I also have another computer Mac which I'm currently accessing through that I'm going to access DJ spark and install Nemo claw there which runs privately and I can even run olo or any local models because this supports large language model running locally because of its graphic card but in our case I'm going to showcase with external providers with nimatron model and you don't need a graphic card for that I have written it as a blog post which I'll put the link in the description for you to follow through I've got all the commands required to run it so as a prerequisite You need Docker installed, Git installed, and curl installed in your local machine. So, I'm going to SSH from a Mac to the DJX box. You can even run on your own computer if you're using a Linux machine because this is fully compatible with an Linux machine, but partially with Mac and Windows. First, I'm SSH into the machine. Now, I'm inside the machine. And then, here are the steps. Step one to 14. Many of the in between steps are just for verification. So first I need Nvidia API key which you can go and create from build.nvia.com. Once after you generate your API key, you can come to your terminal export Nvidia API key like this and then click enter. Next I can combine a few steps together just this command and then running this command as well. As you can see here I'm running those commands. Next installing OpenShell together with these two commands. installing OpenShell, setting the environment variable and that is all done. So first we set up NVD API key next we install openshell and third we are going to install Nemo claw. To do that we need to clone Nemo claw locally. So this is the command for that. So here's the get clone command. Next making the script executable and spark dockerix. So running those two making it executable and the spark. Now the basic configuration is completed. Now we are ready to install Nemo claw. To do that navigating to the folder and then installing this. So navigating into the root folder and then installing. Now you can see Nemo claw is automatically getting installed with all the required packages dependencies. It's able to detect my Nvidia GPU DGX spark using Nvidia endpoints Nemetron 3 because we added our own Nvidia API key that's automatically detected and starting the open shell gateway. Nemo claw endpoint is here creating a sandbox. So the agent will run in the sandbox to be more secure. So this is the gateway inference configuration. It's routing through inference.local. All the API keys or API calls gets routed through this inference.local from the sandbox. It's creating its own policies by default. There are few policies allowed and open claw sandbox is live and Nemo claw installation completed. Now you can see I can open this URL and navigate to that open claw UI. But just because I'm on my Mac and accessing DGX Spock running in my network, I need to run one more command sh and then this command because as you saw the user interface is running in this port. I'm just forwarding that to my local Mac. After this click enter. That's it. And now I have connected my local Mac to my DJX Spark where Nemo Claw is running. Now I need to open the interface. So I can keep this terminal running on my previous terminal. I can just open this URL. So clicking it and here is Nemo claw which has open claw installed. So now I can just say hi and it responded hey I just came online. Who am I? Who are you? So that means it worked. But the issue here is if you go to skills you can't enable these skills because everything is disabled by default. You can't even enable from this user interface. You can see it's missing this missing that even if I try to install clawhub Gemini you can see Gemini came up with an error because it requires allowing this. So how we can enable this? So by default even if any request goes out that is blocked. We need to allow that request. Let's go through step by step. As an example I want a weather. So what is the weather in London? So now the request is going from the chat and it's trying to access details but it's rejected just because we have open shell we can see all the logs and we can allow access. So how can you do that? Let's go back to your terminal there. What I'm going to type openshell term t r m and then click enter. Now here is open shell. Click enter. Now you got the open shell. You can see neo claw is healthy. You can see neo pro. We got the Nvidia API key registered. That's what we are using in prod. And we got sandbox one sandbox my assistant sandbox. So I'm going to open that. Here you can see the networking rules. So for claude api.anthropic.com is enabled for claudhub for discord for github their own endpoints are allowed. So these are default allowed list. So let's see the logs by clicking l. So you can see the logs here. So you can see it's getting denied. So clicking enter you can see it's trying to get the weather from wttr.in and it got denied and it used the curl command. So now we need to allow this request. How can you do that? So clicking escape and again escape and then escape and ctrl c to cancel the screen. You can type openshell policy list my assistant to see the policies which got changed at various time. Now we need to update with the latest policy. To open the configuration I need to open this file that is neo claw blueprint policies openclaw sandbox.yamlo. So I'm using nano to edit it. So here you can see all the configuration which we saw earlier. So this is where all the configuration getting saved. So the read only folder is this open claw can't do anything else such as writing on these folders. It has access to write only in this folders nothing other than that. Then network policies already we saw that all the configuration listed here. So that's where we need to add our O network policy that is allowing to check the weather. So here is the YL to add which allows this website using curl. So going to copy this coming back to my terminal. So just above code I'm going to add that weather allowing this full access and command O to write and command X to exit. Now it's written. Now we need to reset it or assign the policy. To do that we need to copy this command. Setting the policy for that pasting it here and then clicking enter. Now we can see the version number four is getting saved. And that is done. Now if I type openshell policy list my assistant you can see now it's version number four got loaded. Now let's try this again. Go to open claw and I'm going to say try it again. And now you can see it ran the curl command with weather.in and it got the weather data. So that much restricted is this. And in this way your open claw is more secure than ever and it is enterprisegrade and I'm really excited about this. Similarly you can go through these skills run it and immediately you can review that in open shell term in the logs. So you need to type tab to move across and type l to see the logs. So if anything is blocked then you can allow and just now we allowed the weather curl and that you can see here forward allow. So every single thing is logged completely monitored that is the power of Nemo claw. Now once after the fix is applied checking the weather allows and you can clearly see the result. A quick recap on what we saw. We set up Nemo claw in DJ Spark running locally. Accessing the dashboard of DJX Nemo claw from a Mac computer. Adding network policies to allow weather. View the open shell terminal user interface to see the logs. To install skills or plugins, you need to first download into your local machine that is DJ Spark. Once after you're happy going through the files what it does, read and understand. Then you can copy it across to the docker container and then to the sandbox. Then reload the agent. So that's the way you install skills and plugins to open. You can easily switch between AI models. There's a command called open shell provider create. Using that you can configure your preferred model. And I'm really excited about this. If you run any local models using oama then that means like it's 100% free. Everything remains private. So do try and let me know in the comments below what you think about this and do let me know if you have any questions and considering you already like open CL I also create another video about open CL running locally on your computer 100% private I'll put the link in here and I highly recommend for you to watch and I will see you