Transcription
First things first, I want to start with this. Cloudbot is not AGI.
In this video, we're going to talk about Cloudbot or Moldbot or OpenCloud, whatever you want to call it. I've been running it in different instances, different installations. I've tried giving it full permissions. I've locked it down fully. I'm going to give you a very downto-earth, grounded point of view of Clawbot.
There's endless posts online of AI Hypers saying how amazing it is, how it's changed their life, how it's their new always on employee. You may have also heard about Maltbook, their own social network, where they're getting together and talking about things, talking about their humans and creating a secret language that we can't understand. You may have heard it's a huge security nightmare. You may have heard it's running someone's business completely. And not to be that guy, but a lot of it is hype. A lot of it is exaggerated. It can do some cool things and I'm still running it, but let's be honest about what it actually is.
So, I'm going to share with you the nuances that I found, how I've installed it, how I'm running it, what's missing from it, how to install MCP servers, and my honest opinion if you should actually go ahead and spend the time installing it and try to get it to work or just keep working on building your own things.
This video is a partnership with Bright Data. I will be connecting the Bright Data MCP to Cloudbot. And this setup will apply to any MCP server that you want to add to Cloudbot.
Here's the best analogy I could give you for Cloudbot. Imagine having this 10-year-old nephew. He's really smart. He's almost a genius, and for his age, he could do all these amazing things. So, you say, "You know what? Take my credit card. Here are the car keys. Here's a lighter. Here's a key to your own house. Go do whatever you want. I'm gonna tell you to do things for me. And I'm gonna let you figure out how to do it yourself. You have no restrictions, no limits. Oh, and by the way, every time you do something, every time you think, just charge a credit card. Just keep charging the credit card. It's fine." And sometimes he'll do things that are really amazing, really great. You'll be like, "Wow, how did a 10-year-old figure that out?" But because he doesn't understand consequence, because he's gullible, because he's eager to please, it can become very dangerous. Let alone letting a 10-year-old do all these things without supervision is completely insane.
The bottom line is that is what Claudebot is. For all the people hyping CloudBot online, they are giving it way too much access, way too much responsibility, way too much autonomy, and it's just a recipe for disaster. And before we get into it, the whole mold book thing where people are allowing their mold bots to talk on a social network, would you let your 10-year-old intelligent nephew go on social media and make their own language and share secrets? It's just crazy.
With all that being said, it is a capable tool. It's very interesting how it's created and it's not actually that much of a technological marvel. What it is is an agent with unrestricted access that has been hacked together to talk to all these different tools that you can run locally on your machine managed by multiple markdown files and memory files that could install whatever it wants and almost do whatever it wants unless you put guardrails. And sometimes those guardrails aren't even enough.
With all that being said, I installed it several different times. Right now I'm running it on a Mac Mini where I have other users. Turns out you don't have to give it its own computer completely. If you lock it down correctly within its own user, you could have other users running on the computer simultaneously. So I created an entire development environment on one user on this Mac Mini that I could access from my laptop, from my phone, from my iPad. I do that with Tailscale and T-Mox. I'll link to that video above. Then I did chmod 700 to lock down my personal accounts. And then I created a separate user for Cloudbot on the same machine. That's a standard user that doesn't have access to any of my personal files, any of the files of the other users on the computer. It has its own Apple ID, its own email, its own GitHub. Sometimes I'll put it in a shared environment. I'll give it read access. But the number one thing is to make sure that it can't touch your own files. It can't send or do anything on your behalf without your permission.
And I know some people say, well, that's the whole point. Well, no. There's a reason that OpenAI, Anthropic, and Google, and all the other companies haven't released something like this yet. Because every LLM is subject to prompt injections, which means it can't detect, it can't differentiate its user telling it to do something versus a shady prompt hidden on a website or a PDF or a file or anything. This is a real problem that LLM have. And that's why for the most part, these companies haven't released something that has this full autonomous ability cuz there are people in the world that will convince it to send your private data, send your credit card information, send your API keys, whatever it is. People have tried it and a gullible 10-year-old or an agent that has unlimited access without guardrails will fall for it.
So the setup I recommend is on a separate computer. It doesn't have to be a Mac Mini. The idea is that it's always running. So once you've created a new user, you log into that user. You don't connect your Apple ID. You don't connect your Gmail. You don't connect anything. You could create new accounts for it later. You'll install it by following this command and it has a CLI that will walk you through how to do it.
Now, keep in mind because I installed Homebrew on my main user, it actually adds this extra guardrail that stops Cloudbot's ability to go install things on its own or just makes it a bit harder. For example, a lot of programs that will install on Mac are via Homebrew. Claudebot is not able to use Homebrew. So, if it wants to install something, it has to tell me, I go to my main user, I install it via Homebrew, and then it becomes accessible for Cloudbot. So, it's like if your 10-year-old nephew wants to go get a tattoo, he has to come back to you for permission. He's limited in that moment. And by the way, this isn't foolproof. There are other package managers it could still use, but it does tighten the leash a little bit.
So, if you're installing this on your own machine and you're going through the configuration, it's important to choose loop back, meaning it's not opening a port to the rest of the internet, which is one of the biggest concerns that everyone's talking about where everyone's cloudbot is just accessible to everybody else. Second, don't install every skill out of the box. I recommend a very vanilla, a very limited setup. Again, yes, there's people saying that it's doing all these amazing things. Those amazing things cost money. Those amazing things have API keys. Those amazing things have vulnerabilities. So, I recommend the most limited version of CloudBot. At least to get started, see how it works.
And after running this for over a week, I'm convinced that a lot of these things were already possible with things like Cloud Code. Maybe not out of the box. Maybe you had to do a little bit more work. But at the end of the day, it's more controllable and more efficient because I've noticed two main things about running Cloudbot.
First of all, memory efficiency. Yes, I'm running a Mac Mini. Not the most powerful Mac Mini. Definitely more powerful than an instance you run on the cloud. But I've noticed memory spikes when doing tasks cuz everything is managed by multiple markdown files and it will keep reading and writing. And I've seen this almost brand new computer grind to a halt when ClaBbot starts doing things. And when I dug a little bit deeper, I noticed it had to do with its architecture. And yes, this is an early version. This is definitely not a production-ready tool, but it's just interesting to see.
Two, it's expensive cuz even though it's open source, you're running this via a different Frontier LLM and I've tried a bunch of them. The bottom line is if you want it to be effective, you want it to be good, you're probably going to want to use Chad GBT or Claude. I've tried the smaller models. I've tried running it locally. I've tried the Chinese models. GBT 5.2 or Opus 4.5 are the models you're going to want. Anything else just doesn't work that well. So, I decided I'm only going to run GBT 5.2 cuz they have better usage limits. And the truth is, it was just boring and it just wasn't as good. It worked fine. It just didn't feel as good to me. So, eventually I connected it to my Cloud Max plan. And here's the crazy part. Friday mornings, my usage limits reset. And yes, I'm using cloud in different instances, but within five hours, I was already at 10% of my weekly usage. So, not only is it expensive, it runs your usage limits up so fast.
And for those that will say, you know what, I'm going to run local models on it. Well, my friend, good luck. You need a very, very, very powerful computer to run the models that are comparable to GBT 5 or Opus 45 to get an experience anything close to these Frontier models. So, that's just the truth about efficiency.
So, when you're going through the installation process on your Mac, there's a few things you should do. First, you're doing local. This machine, no one want remote in web tools. It initially only gives you Brave search and it gives you the ability to turn on web fetch. Both of those, in my opinion, are not good enough for an agent that's supposed to do actual work for you. If you want this to actually make decisions and get context, you need a better tool to search the web. And that's where Bright Data comes in because they have all these built-in data sets and they have a scraping browser. We'll get to that in a second cuz we have to add that manually as MCB server for the gateway.
If you're setting it up like me locally on a computer and you follow the next step, this will be fine because the next step is gateway bind mode. What we're going to do is loop back. That means the gateway is local. It's stay on your computer. No one else can access it.
Channels is the way you will interact with Cloudbot when you're not at your computer. In other instances, I've tried iMessage, I've tried Slack, I've tried WhatsApp, and I've tried Telegram. And in my experience, Telegram is the best one. It's easiest to set up. I'm not sure why it says plug-in disabled here. That's something else I want to tell you about. The configuration here and even the gateway don't always make sense, but we'll get to that later. The bottom line is this is the external channel you'll use to interact with Cloudbot.
Okay, skills. Now, there's a few ways to install skills. I'm using mpm and these are the skills that offers out of the box. Now, you have to install the dependencies for the skills. But I'm going to tell you, don't install too many skills in the beginning. The only skill I recommend installing is MC Porter. That's how we connect MCP servers to Cloudbot. And the good news is this CLI will walk you through the whole entire setup process.
So once you get it all set up, it will give you a URL where you could access the Cloudbot gateway. Here's where you can chat directly with your Cloudbot. Obviously, you can chat with it with the channel you set up if it's Telegram, WhatsApp, whatever. And you can still interact with it via the TUI. You can do Cloudbot TUI and then you'll be able to chat with it in a terminal.
Cloudbot manages sessions and memory in these weird ways. You can actually check it all out. You go to your user directory and you open up Claude and you can see all the files it creates and how it organizes everything. Now, in the gateway, you can see all these things. The cron jobs you create that's like the reoccurring tasks you can configure a lot here.
Now how do you connect MCB servers? So the way we use MCP are in skills here. You see a list of skills and when it means blocked it means that the dependencies just aren't downloaded. It's not actually blocked. Sometimes disable means enable. It's really weird. So you can see all these skills that are built in the same list we saw before when we started config.
To get MCP working, you have to install MC Porter. So you press install MC Porter. It will install. You're going to have to restart the gateway. Then we go back to the home folder where cloud exists and we go into this config folder and here we'll see the mcporter.json JSON. But first, we have to get our API key and our configuration.
So, we're going to go to Bright Data. If you don't have an account already, you could create an account. I'm just going to log in. I'll drop a link to Bright Data in the description below. And if you click that link, you should get some extra free usage out of it. So, once you're in, you go into the MPCP. For those that don't know, Bright Data has about 63 different tools, a lot of really great tools like searching Amazon, Walmart, LinkedIn person, LinkedIn company. It's really good at searching Reddit. It has its own scraping browser. It could look at a GitHub repo. There are so many use cases for it. Sponsorship aside, this is the only MCP that I install in every single client because it's able to access the whole web.
You actually can select which tools you want from the config. Then it will add it to your MCP config string with your API key. So check this out. Let's say we just wanted to give it e-commerce and financial intelligence. You click those two. You press continue to configure and it gives you the exact string you need with your API key. You should definitely use streamable HTTP. And then depending on the client you're using, it will tell you exactly how to add it. In my case, I'm just copying this whole string and we're going to go back to the MCP portter config.
Now, what I'm going to do is just open this up in VS Code. So, when you open it up, it looks empty like this. All you have to do is give the name of the MCP server. In this case, Bright Data base URL. Give the exact URL it gave you with your API key and all the tools you selected. And that's it. But you don't even have to do that cuz once you have MCPorter installed, you can just tell Cloudbot, I installed MC Porter, here's the remote URL for the Bright Data MCP server. And it will add itself. And this should work for any MCP server you have.
So we can say use the bright data MCP to look up Jared Blumenfeld on LinkedIn. And there we go. Here's what Bright Data pulled for Jared Blumenfeld on LinkedIn. So now we see that the Bright Data MCP server is working.
Now the one other thing I should tell you about connecting MCP server is MCB Porter just exposes all the tools of MCB server. But if you want it to work well, I suggest you explain to Cloudbot what the MCP server is, the tools it has, and to always use it when you want to perform a certain task. Once it understands that, it will create a host skill, save it to its skills directory, and it will be able to keep using that moving forward.
By the way, these are just the skills they offer you when you install Cloudbot. There's a website called Claw Hub with a bunch of other skills that people have made. So, you can install it to your Cloudbot. And just like how I recommend vetting your MCB servers and vetting your skills, these need to be vetted even more because they don't have the guardrails. They don't have the safety that Claude or Chad JBT have built into them.
Now, as you're soldering the setup, it doesn't always work. It looks like it's working. Sometimes it looks like it's not working. It's not so stable. And after using Cloudbot for over a week, I've realized that Cloudbot is cool, but the security and efficiency questions, the time it takes to set it up, the time it takes to make sure it keeps working. This is a fun, cool application, but it's half-baked and there's a lot of open questions here. I think if you want to tinker around and do some cool things, go for it. But all these videos online showing all it can do don't really give you the whole picture. And I think if you're trying to build stuff, you're much better off just building things with cloud code or codec or open code.
In any case, I'm still running Claudebot. By the way, after editing this video, I just saw that Bright Data added their own skill to Clawhub. So, if you don't want to go the MCP route, you could also do this. I'll drop a link to this below as well.
So, I hope you found this video helpful or insightful. If you have any questions or feedback, drop it in the comments below. Thanks again to Bright Data for sponsoring this video. And if you haven't done so already, subscribe to the channel. It really helps me grow. Thank you guys for watching and have a great.