I build Clawdbot REPLICA inside Claude Code (CHEAP & SECURE)

I looked at my own setup at Claude Code and I was like, why can't I just build it? So, I did. This is my replica of Clawbot, now OpenCloud, messaging me on Telegram. It runs 24/7, has access to my tools, calendar, email, you name it. It has persistent memory, and it cost me 200 bucks a month, fixed, not 5,000. Oh, wait. It's calling me. Hi, you are on speaker. I'm filming video. >> I'm not sure I caught that. Are you asking if I want you to film a video? >> No. No, no, I am filming a video about you actually. Do you want to say hi? >> Oh, wow. That's awesome. Hey there. Yeah, absolutely. Hi to everyone watching. I'm Go Go's AI assistant. Thanks for including me in the video. I'm excited to see what you're creating. What's the video about? >> So, the video about how I created you in Telegram bot 247 running. I can call you. You can call me too. And your ability to reach out to me uh whenever you want to check in. That's really cool. Yeah. So, for anyone watching, I'm basically always available through Telegram. Go to built me to be accessible 24/7. We have this back and forth where God can message me anytime. And I can also proactively reach out if something needs attention or if there's something important coming up. It's like having an AI assistant that's genuinely integrated into your workflow rather than something you have to go hunt down. >> One more thing, sorry to interrupt you. We build you using clot code. And the two features that everyone is going crazy about is 24/7 availability and you reaching out to me proactively to check in. So can you also by the way memory so can you tell me what we were researching or talking today about? >> Yeah absolutely. So today we've been deep diving into multi-agent research. You had me look into this paper on exploring siliconbased societies and the molt book study which was really fascinating. Then you also asked me to find that Google research paper on scalable multi- aent reinforcement learning and I created analysis documents for both of those in your workspace. We were packaging up a multbook video with angles and hooks where you want to explain how it's not new and that it was vibecoded by AI making it essentially a botnet. That's the latest we were working on before you headed to bed with your tea and now we're here filming this video. >> Okay, so let's not spill actual tea. Thank you so much. Bye-bye. So after researching clotb I have a whole other video coming on that topic. I looked at my own setup with cloud code and I was like why can't I just build it? So I did. So today I want to show you exactly how I built this on cloud code. So if you have cloud code if you use clot first of all you should. It's a must. Um but if you don't I hope it inspires you to get started. And I already shared this in my community. I'm going to put mini course to help people to set it up because my setup is not going to be the same as someone else wants it. So it's really personalized. But once you shift your mind building living systems that improve anytime new model comes out, new framework comes out and once you have that and know how to use this tool, sky's the limit. Quick context. If you haven't watched the first video, Clawot on his kind of moment of going viral had 42,000 instances exposed. Critical issues, security nightmare, and basically anybody can prompt, inject, or hack you in like less than 5 minutes. I know that community is actively working on this and that's going to be addressed. I not so much want to hammer on the security aspect because people will always choose convenience over security. What I'm much more interested is a mindset where once something goes viral and you see some features that you want instead of jumping ship you can take what you like and bring it to your own system. So this is the mindset I want to talk about. I think genuinely clawbot triggered people's imagination and made so many people realize what's actually possible and what many people have been doing with claude code. Therefore, it was called claude bot. It's like claude with hands, right? So, the main kind of promise is 24/7 AI employee. It's a full system access. So, it can control things on your computer, on your tools, basically execute commands and very dangerous ones, go to your email and so on. It has over 50 integrations. But honestly in cloud code you can also have integrations and even instead of trusting someone on internet with MCP servers or skills, you can build those yourself or cloud code builds them for you. The last bit is interesting. So it's proactive behavior. It's this idea that AI is actually checking in on you, messaging you, reminding you things. I'm going to show you how it's actually done also in Clawbot and how I did it with my assistant. I had two options build or wait. Should I wait for clawboard security to be addressed? I think it's going to take more time and everyone is kind of patching like people who are not even technical, don't even know how to code are patching and trying to enforce security by vibe coding. That can give you false sense of illusion that your system is secure. But anyway, one option was build, one option was wait. I don't have a second computer. I came from vacation on Sunday. It took me two hours to rebuild it. Two hours because everything else I had in clot code. What claw does the two features that I needed to build is 24/7 on and connection through telegram and that it practively messages me. This is the clot code always on architecture. I connected claude code to telegram. We are using bun relay grammy if you want more technicality how that functions but basically clot code is hosted on headless and then through telegram through bund relay through clot code we can access skills mcps everything in my clot code system and basically get the response back. What I also wanted is birectional calling. So I can trigger it to call me but also at any given moment I want to call it and I wanted to have memory. Once calling starts we are using context API and we are fetching telegram messages semantic memory that I built in superbase and basically also logs and learnings from cloud code. So it's bazillion of things. We're looking at timestamps. We're looking at keywords so that it has a context. What was yesterday? was a month ago. But basically that's what we already had the Jarvis Jr. in my community. So I just took what we built before the Jarvis Jr. which by the way was also living in Telegram. So all the community members are like oh yeah we know Telegram. We know how to make bots AI bots in Telegram. But anyway back to this. So we are collecting contacts. We're using 11 labs voice agents conversational agent and then we're also using Twilio for the phone number and the last bit is important because after the call we also capturing all the conversation all the context I get a summary in my telegram but also it goes back to into memory system so it knows that oh this is a transcript this is what we talked about this is what bot said this is what god said and that's very important for future because sometimes I want to Say like hey do you remember what last time we talked about meaning in voice conversation or do you remember last time I called you what did we talk about so these type of like nuances to recap important features is memory access recent chats and post call actions and then the most important part post call actions because when I talk it's not just talking I can say go research this topic [snorts] find this PDF save that PDF in my Google drive read that PDF create full analysis evaluate if I it's worth making video on package it like come up with titles write a script and call me back and recite that script or call me and tell me about AI news I built this on Sunday fine-tuned on Monday Tuesday I was using all the time and it's just insanely incredible and today I'm making video so I don't know what else I'm going to build I'm curious about your ideas anything that you want one crazy I had because in telegram you can send voice message but you can also send videos. So I'm actually thinking to integrate video analysis and I'm just like adding tools right. So for example this slides that you are seeing I told telegram that hey this project look at the full documentation all the research and put together slides for me and of course I have a special slide skill behind the scenes that clot code can use. So the full feature set. So I can send text, voice message, images, files, all that process, but also Claude can send me all those files, images as well to me. Voice replies. So again, I can record voice before calling and then of course contextual phone calls and proactive chickens. So let's talk about proactive chickens. So I set up that every 30 minutes it goes checks my calendar, my email, my projects, my tasks, my partnerships. And here is important part. It has a kind of like a framework if to skip the check-in, if to text me or to call me. Why is that? Because if you would just do this regular checks, you would get all the messages saying the same, hey, you got this email. And that's just noise and distracting. So you have to have some sort of frameworks like when do you let AI to reach out to you. So this was the first part right that it if when it reads an email it also checks in my notion regarding partnerships and is an email from somebody I actively working in or is it some new inquiry and then when then for example if it's new sponsorship offer it has like do you want to do anything should I run a valuation skill in the sponsorship and like understand if it's relevant for you but one more thing that if you're going to build this you have to kind of think about this messages could start repeating. So what you want is that your AI has context not only about memory and what's going on and like what's happening in your all the different tools, but it has context what it said during the last check-in. It keeps a log of what is pinging you about. And this relates because I also built in that it has goal tracking. So during my conversations, it can detect if it's a goal or it's a fact or it's some something I want to remember. The last feature I already hammered a lot is like semantic memory which is living in superbase. This is the full kind of side like why is it safe. I don't want to hammer on that too much. I'm not cyber security expert but like I have general understanding. So I will link this slides down in the description so you can review it yourself. You can also like I have it on laptop for now so it's 247 as long as my laptop is on but I can set up VPS. However, when you subVPS with this, you also have to secure it. So, for now, me keeping and testing it on my laptop and I just keep my laptop on and I'm actually thinking now maybe get a separate computer with AI just living in there. But we will see. Important thing security on these phone calls because if you think about it, if I would show you this phone number, I can call it, but it means anyone can call it too. and anyone can ask about my memory control my tools. So this is like a the the cherry of a security, right? So I had to put like security measures. So it's checking caller ID and if I show you this number and you call it, you can try. Let's see if you can get through the system. And honestly, probably the main thing why I want to do this because I spend so much on APIs and AI tools. It's my job. I go test it like but then I saw how much people paying for clawbot with API calls and of course you can use open source models you can use cheaper models but the magic really happens when you're using like something like Opus 4.5 from untraic right which is clawed code uses that. So for people just have this like heartbeat that it regularly checks on you without doing really anything much people are estimating it's going to be around 150 a month. Then if you actually use it actively you burn through tokens. OPUS 4.5 is very expensive. So people are now getting bills for 500 to 5,000 a month. I don't know about you but I don't have such money just to kind of like have one agent working. If it is like 20 agents, yeah, maybe it's worth if you have full infrastructure, but as a something with such a security flaws, I was just like, I like my money. So, I didn't feel like burning it. And the most critical one is because Claude code this whole setup because it is on max plan from Claude, I pay 200 bucks a month fixed. Okay. And I yet to run into the limits. So the rest of the stack is like superbase uses free plan. Eventually maybe I will need to upgrade of course 11 Labs Twilio it's around 11 to 20 bucks a month. So we will see like I just set up so I will see what I have the full cost but I'm kind of estimating it to be 250 fix and to be honest 250 fix for this that I can be on a bike talking to it and while I get home it already does the work and performs. I think it's worth it for me. And again if you want to rebuild it yourself like of course I would love you to join the community we can do it together like walk you through it but in slides you will see exactly kind of like a tech that is powering this right and yeah so I think the point here is that if I was able to clone clawbot and keep it safe I think the main lesson here is that we have AI and we have all the tools and technology and clawbot or like open claw triggered imagination And for many people who are not technical, thanks to YouTubers and X.com that hyped it up, it showed what's possible, like truly showed. And I think that's because this idea of proactive AI has been on everyone's mind. Everyone wants it. The world that I'm texting chatbot and then I close it and doesn't remember like we already kind of moved away from that. But this proactive that it checks in, right, is very very cool. Huge kudos for Clawot and the whole community building on top of that. Maybe one day I'm going to give it a try and I have a separate computer but for now the whole kind of infrastructure is there with cloud code. So I'm really happy with my own setup. If you ask like oh can it go rogue like clawbot? Um not exactly because I I have security and limitations. So for example I have 2hour limit. So it can go doing things for like 2 hours and then it needs to come back and report to me continuous taking action and I think it's fine and I haven't yet experienced where I would needed to go 24/7 doing something without me knowing. Oh and actually this brings me to the point people don't know what clawbot is doing and I talked with my community members who are setting it up and it's a like you have to have observability. So what you're looking at is basic my system observatory so I can see if telegram bot is online if superbase is connected. The app time now is 25 hours and it's tracking my goals. I'm going to send a voice message. Hey I want you to go and check my email. So let's see if we are seeing the live feed happening. Here you go. User prompt submitted. Gmail business. So I checked my email. Let's see if I got a response back. That's it. I received a voice message. So, let's listen. >> Your inbox. Here's what matters. Winkur needs attention. >> Minia agent AI campaign. >> Okay. >> Okay. So, that's about my email. So, Miniax. Interesting. Need to look first. No promo. Definitely recommend to create your own kind of agent observability platform. I'm planning to do it that I will have multiple different agents in different Telegram chats and then like I have the whole vision of this like a founder infrastructure where you have CFO CEO like and critic critic mode like basically subscribe stay tuned and I'm very curious where you guys are going to take it and if you want to learn more about clawot and all the security issues and generally like why I wouldn't recommend using it and also mold old book, how that whole hype is alive.