Living with AI Agents: OpenClaw & NanoClaw in Practice - JuniorDevSG

Hi everyone. Thanks for coming tonight. Um yeah, it's on a Friday night. Hopefully everyone is excited for the weekend. [laughter] >> I guess only one. >> I mean I have to work this weekend so also not super excited. >> Okay. Okay. So uh just a little bit about woman dev before we get started. So we are a woman in tech community. I think we can say that we're one of the more most active woman oriented communities in Singapore right now and we are formerly known as woman who code Singapore but in 2024 that got bankrupt. The parent organization declared bankrupt. So we had to rebrand as woman devs. So that's just a little bit about oursel and we have events ranging from technical events, workshops, panel discussions, things like that. So very similar like junior devs but our audience is mainly for women. So yeah, that's about us. So who should join? I think like I said, our community is meant for professional women who who trying to find a place in tech. Uh it can be also mid-career switchers or anyone who is a non-developer but just interested in tech in general. They can come visit our meetup. It's on our link tree here and they can come and you know check out all sorts of events we have. So we have talks just like this for tonight. So yeah. Okay. [laughter] Okay. Uh so this is our upcoming events. So we have a pitch night on uh August around oh sorry September September 11. Sorry my brain but yeah we have a event coming up pitch night. So it's where people come in and actually pitch about their product ideas because we noticed that people in the community they always build like site projects but they don't have a platform where they can actually showcase them and receive feedback. So we have this ongoing. So feel free to register via the QR code. Yeah. And then another another one is the pitch ready workshop. So that one is actually to help the speakers in the pitch night to actually uh get ready on like you know learn the skills on how to pitch your products. So if that's something that Yeah. Yeah. This is why I remember 21 August. Who's the pitch ready one? Yeah. So the pitch ready one is where we actually invite uh an author to come and actually teach you the skills on how to get ready for a pitch which will be on September 11. So yeah, if this is interesting to you or if you know a female colleague who always wanted to have the opportunity to share, feel free to share that with them. Yeah, another of our upcoming events is we have the uh public sector tech early careers fair. So that one is happening at September 6. Yeah, September 6. So feel free to register here as well. So these are this is a career fair where you can connect with people and yeah hopefully career. Oh and we have a Hectober fest upcoming. Sorry the poster is not really up to date. That was from last year but yeah this year woman devs is also participating in Hectoberfest. And what Htoberfest is, in case for people who don't know, Htoberfest is a one month long opensource uh kind of campaign that's going on by Digital Ocean. So every month from October 1st to October 31st will be Hacktoberfest. So Woman Dev, we do have open source projects in our GitHub repo where we will actually uh encourage a lot of people to uh help us with and contribute to our open source project during that time. So if you're interested, feel free to check out our GitHub repo at githubsg for our open source project. All right. Um this is just my shameless plug. Uh I also run a podcast called React Tag. So yeah, one of my co-hosts Looney is also here. >> I don't want to be the only one like in the spotlight, but yeah. So me, Salony, and another member of Women Dance, we actually started a podcast because we realized that we always yap about things around tech. So might as well record that and turn it into a podcast. So yeah, now we have a podcast and feel free to check us out. We talk about things that's happening in tech from a Southeast Asia perspective because we noticed that a lot of the media in tech is so Silicon Valley and so Western audience oriented and we just want to give like a local Singaporean perspective on that. So yeah, feel free to check it out. We also have guests like for example Sudier. He's one of our friends who recently published a book. So he came on our podcast and we talk about the global south about AI in the global south. Yep. Okay. So now we get to the real stuff. So now is um this is the talk me and Zawana have done actually for woman devs uh a few months ago but it was very popular among women because at the time it's our woman devs community. So not a lot of um men showed up and then I guess we want to open it up for more people to also enjoy. So that's why we're doing this talk again. It's basically about living with AI agents um about my journey in openclaw and her journey in nanocaw. So first a question for the audience. How many of you have an AI agent? Oh okay. Okay. Oh not not a lot. How many have like like personal AI agent working for you like 247? Ah okay. Okay. What what do you use? May I ask >> Burmese? Oh, that that's a good one. And your >> Sorry. >> Email. >> Oh, email. >> Email. >> Oh, >> okay. [laughter] >> Okay. Okay. Okay. Yeah. So, hopefully today you'll learn a bit about open call. And actually, I I do want to explore Hermes as well. So, maybe we can talk after that. Oh, yeah. A quick introduction about me because I've been yapping and you don't know who I am. Um, my name is Victoria. So I'm a solutions engineer lead. That's my daytime job. But my but I have a lot of jobs besides that. So I have a 5 to9 after my 9 to5. So my I know I know I'm busy. So so yeah. So my 5 to9 is I'm also a technical blogger. So I have a blog at low-fvictori.com. Feel free to check it out where I also write about my journey in you know openclaw there. I'm also a co-director woman devs. So like I mentioned woman devs the community for women. And then I'm also a co-founder at Red Tag, which is the podcast I just shamelessly talked about just now. So, so yeah, if you think I'm a workaholic, don't worry. I have hobbies. So, just just in case you think I only work, I I play games. I play video games. So, if you want to come talk to me about games or books or tea, anything like that, feel free to also talk to me. Oh. Oh, yeah. And this is just a little a more shameless brag, but um recently our podcast we won a award for creative use of social media. So yeah, this is me Floonian. Yeah. So yeah, thank you. Yeah, thank you. Oh, thank you. So yeah, this uh the second award we won this year. So yeah, our podcast is doing well. Okay, so now uh finally, I'm sorry I've been so VIP, but now it's about me and open call. So before I talk to you about how I get into open call um I mean sorry before I talk to you about my journey with open call let me talk to you about how it started. So basically we attended an open call meetup in April right? How many of you have attended the open call meetup? Okay. Okay. Okay. Okay. [laughter] Okay. Not not the majority, but during the open call meetup, it was like early April. So that was when open call is still kind of in that stage where it's not super um the most trusted agent framework because it does have security issues and all that. So after I attended that open claw meetup with me and my sister h we decided to actually explore open call just the possibilities because we realized that yes it's not the most secure for non-developers but we are developers and we know how to enforce those guard rails and so therefore I went and realized I have a lot of big plans. So this is my sister. This is just a chat with me and my sister. I just say I I have a big plans now because now I see that the possibility of creating open claw is my own AI agent, my own secretary 247 work for me. So yeah and it was it was kind of a high I would say I'll describe it like it was a you know when you're like so into the flow state you get super high. So when I was researching about open call it felt that way. It just feel felt super high and just us putting that plan into action. >> Yeah. Yeah. I I'm sure you guys know. So, but unfortunately I was trying to integrate Google my open call to work with Google but Google actually detected that is a bot. So, Google disabled my my open call within like a few hours of it being created and connected to my uh my bot. So yeah, I call my agent Zenbot. Yeah, it's just a name because it's from Overwatch and my main in Overwatch, which is the game, is Zenyatta. So Zenbot. So yeah, I I told my sister I'm breaking up with Zenbot because it's taking too much of my day and like I was like crazy addicted to Zenbot. Like every day talk to it like yeah, it was it was addiction. It was very bad for me. So in a way when Google banned my bot account, I thought, you know what, it's fine. my vision won't be realized but it will just be somewhat of a blessing in disguise. I try to see things in a positive light, right? But oh well, I appealed to Google and Google was like, "Well, we shall restore your account." So, it's actually not such a long breakup after all. It's just a one day break. So, my Zenbot is back as you can see. So, yeah, it's very good. And I end up expanding upon that and which you will see later in the talk how I actually built Zenbot now and how it does almost everything for me. And yeah, that's just a me story about how how I do it. So yeah, uh just a quick breakdown of some other personal agents out there. So actually a personal AI agent, how I define it, it it has a brain, it has hands, which is the tool execution, it has the memory, so it knows who you are, what you do, context for you, and it has a heartbeat, which is proactive autonomy. So things like the clown scheduling, things like checking in on you, it can do that. So instead of like an LLM which like you always have to prompt it, it can actually prompt you, right? So that is what I define a personal AI agent like. So I think someone in the audience mentioned Hermes. Yeah. So there's there's nano claw, open claw and zero claw. Yeah. How many of you know about this or heard about this before? No. Okay. Okay. Okay. Yeah. Yeah. So yeah, this is these are personal AI agents. So these are things that you can build for yourself. So hopefully after this talk you might feel inspired to build it for yourself. Okay. But what about these? How many of you use these more people? Yeah. Right. So what what do you think are the main difference between these and these? Does anyone know >> from you? >> Oh yeah. Okay. Kind of leaked the answer just now. Okay. Yeah. Just to save time I'll just give you the answer. Yeah. So for a personal AI agent, yes, like you mentioned, it has context, right? So it remembers things about you and it can plan and use tools. So it's like an agent, right? It is connected to the tools and you can give it skills that can use to connect to these tools. Next, it can be follow up, you know, it can prompt you, ask you questions. So I would say an AI agent is more about goals, right? You need to know what you want from it. So give it goals to do and you design around that and then it can perform. Whereas for the usual chatbot that we talked to that one is just generation right text gen video gen things like that. So it's waiting for you to prompt it to give it an output. So that's the primary difference. Yeah. And how about these? Does anyone know what these are? >> Yeah. Okay. >> Coork. Yes. Uh, last one. Last one. No. Yeah, it's a bit not that not out in the news so much. Scout. Scout. Microsoft Scout. Yeah. Okay. So, yeah. Basically, the difference between these and these are they're both AI agents per se, right? They have connections to tools. They can connect to tools and they can uh run things for you. But the net the biggest difference is this one is self-hosted and this one is by a company, right? So for example for clot co-work you have to use it within cloud's ecosystem. So you have to use their LLMs you have to use their way of doing things. Whereas for open claw like which you will see how I use it later. It's totally free like in terms of how you want to self-host it how you want to design the your tools around it. It's completely autonomous and you have complete power to use it. So you can say that there's some trade-offs here. So let's say you want to automate your life, right? You have to think about these trade-offs that I put here. I think that's the main thing that you have to think about when you're choosing because it does make a huge difference when you're trying to automate that. Yeah. So now just a quick brief about open claw. So it is a operating system for agents that that's how I like to define it. It started from a hackathon. Actually the founder of open claw actually wanted to make a way where they can have a they have WhatsApp connected to cloud code and you can just chat from clot code uh using WhatsApp like that to text uh and it just name it cloudbot right because cloud code so clotbot so at first that was the name of our open clock is cloudbot but of course there were some legal issues and things like that so it becomes open clock now yeah so fast forward it becomes open call So yeah, that's a bit about open core. So now just a quick brief about the anatomy of open claw. So for my open claw at least, I decided to self-host it in my Oracle cloud because I'm very cheap. So Oracle cloud actually has a free tier in case you guys don't know. You should check it out is it's entirely free. I don't have to pay a single scent so far. And I've been using Open Cloud for like I know since April. So how many? Yeah, four months. A little more than four months. And it's completely free. So do check it out and shameless plug I do write a full tutorial about it in my blog. So you can check out my blog low-fvictori.com. So that's how I set up my open claw with Oracle cloud. So yeah, it's very good with the free tier. You get access to yubuntu, you get access to the 6 GB RAM which is more than enough actually and you get you just connect it via SSH. So yeah, it's very good. It's 247 zero cost. Amazing. Um yeah so another thing about open claw that you should know is that open claw starts with a few markdown files when you first spin it up it has the agent MD. So this markdown file is where you define its behavior and instructions. This is where you talk about the goals that you want for it. For example uh what kind of agent are you? Are you like a personal secretary? Are you like a scheduling assistant? So give it an identity essentially. So this is where you can write it and of course you can always update it over time but yeah it generally is where you define that. Next is the soul soul.md. I I know I don't make up these. This is how the the creator of open claw may make make the thing. It sounds like it's a real person right with the soul and things like that. So this is where you can define his tone like how you you want it to speak to you in case you're very particular on that. All right. I mean I I'm not really but but some people maybe they want the soul. So yeah you can say oh talk to me in a pirate way or something or British way I don't know you can define it there. Next is the tools. Oh so this tools you will use it very often because over time as you build on your open claw you will build more and more tools for it right makes sense. So you will have to always say to open claw remember this in your tools MD because this is where it will remember it's like it's dictionary of all the tools that it can access. So this is very good. Next is the heartbeat MD. So, like I said, it actually helps you. It can actually checks in on you. So, you can have this in your heartbeat MD where it can cons consistently do background checks for you in case you need it to maybe, I don't know, check on your workout. [laughter] I don't know. But you can you can do that. And that's how it actually runs the chrome jobs as well. It actually checks whether or not there's a chrome job that's supposed to be run. So, that's the heartbeat. And finally, is the memory. This is very important because again without memory it's just another super power chatbot. I know it sounds good already but with memory it sounds a super super power chatbot because with memory then it knows exactly who you are. Like I can just ask my Zenbot now like what am I doing tonight and for what context, right? It can fill in the blanks for me because it already knows what I do in my spare time things like that. like not like spare spare time but my five to nine not everything I have to tell him right or her or they I don't know but anyway uh yeah so how you connect the open call is through a chat app right so for open core you can connect it to oh I don't know over 45 different chat apps like discords telegram so for me I use telegram so it has a websocket server that it handles all the incoming messages and then you can always you know that's how it runs runs every request through the agent runtime and there therefore the user interface and the execution is separated right so anything you talk to the Zenbot or the open call agent it actually runs it via its own framework and that's how it actually executes your command one at a time so not paral parallelly sorry I cannot pronounce the parallelly it pronounce it one it does it one at a time yeah okay so Yeah, overall this is a flow just an example flow on how the open call works. So first I send it a message through telegram bot right so I create a telegram bot for it and then which I call it as zenbot. So of course you have to do all the botfathers things in case you don't know you can always look it up on Google how you set up your own bot and then you connect it via the telegram API and then again like I mentioned the web hooks and then it go to the open cl gateway and then the agent runtime will parse the intent and they will actually try to make sense of what you're asking it to do. If there's any skills or tools that it should access it will do that as well and then it'll draft a response back to you. So that's the high level way of how it connects. But now this is my favorite part which is all the stuff I've done and this was back in April I think it has been updated now. So but this is just uh in case you are thinking of use cases for your bot and oh Victor I no idea this your ideas. Yeah. So these are things that I've done with my blog I mean with my bot and there's more. These are just workflows. So these are just u things that I wanted to do every day like a cron job but there are tools also. So I have another I have another another set which are tools. So these are things that I'm doing now and there's like more upcoming upcoming. So so yeah again if you don't have much ideas maybe this helps. I don't know. It's it's very customized for me though like see Japanese vocabulary tracker is so random but I'm learning Japanese right. So this [laughter] so yeah it's very personal right that's why I say it's it's something you can 100% customize it's totally up to you that's that's the beauty of having this self-hosted DIY agent all right so quickly I tell you some of the good skills that you can start let's say you want to start open code today this is these are some of the skills that is right off the box that I recommend you should download I I install which is gogg the Google workspace it's so important that the bot can connect to your Google. So I create a dedicated bot account so it doesn't affect my personal ones but it has a bot account that can read my calendars. I have four different calendars, right? My woman desk, my rect tag, my personal and my other bot. So I I have four calendars and then this one Zenbot calendar can read all of them. Yeah. So that's how it can give me a daily briefing like oh these are all your schedules for the day based on all these four different calendars. So yeah, connecting to Google is a must is really helps your life but make sure you know you always build with security in mind right so don't give it full access like it can read my email but I can't send any email things like that yeah so I give it very limited access just for my own peace of mind uh yeah next is the nano PDF so I have a lot of PDF that I have to do in my daily work and I'm so lazy to like read so I can ask the Zenbot to summarize all these PDFs right or even save all these to memory so it have context for future like it can become a rag bot in the future I can ask it questions from a printed PDF and it remembers so again very useful to have the nano PDF as a skill next is the selfimproving agent so this one is the founder of open call himself make this skill because he thinks it's important that the agent itself always keep improving and always get to know who you are and what you do and what you can do to be better so This is the self-improving agent. It's a good one. Next is the claude docs. So again, the founder at the time is called claude bot. So they didn't rename it. Still called claude docs. So this is the documentation skill. This is where it can look up things that you know it should know as a open core and you know things that it should do as best practice. So it's good that you can give it this skill as well. And then finally, GitHub. So I have a Chrome job daily at 4 a.m. where I wanted to back up all its memory because it memory is so important, right? Like I mentioned. So I have a GitHub repo that is private and it's my open call backup repo. So in case something happened to Oracle suddenly it's like oh sorry Victoria is no longer free and I'm cheap right? I still want free stuff. So I will have that backup and move it to somewhere else right. So always have a backup of the memory of all the workspace and all the tools that you have built with your open claw on some you know version control like GitHub for example. So I use GitHub and that therefore I downloaded skill. Yep. So so some of the use case that I talked about earlier like it's one of my 22 workflows. I just want to give a few demo. So this is the daily calendar sync. So again, I have four calendars and my one central bot will tell me what I have to do for the day. So for example, this is back then in April. Yeah, long day. So there you go. This is the 8:00 a.m. every day. It gives me this. And then next, the to-do is. So I I have to-dos also. I I have things to do. So it also give me things to do. It'll tell me. So I can, you know, add task. If I wanted to add task, I can ask it to help me um remind me things from the to-do list. So basically, I just give it the API key for to-doist and it can do all that. Yep. So what maybe some of you ask, oh why not Google task? Why why to-d doist? Will you guys ask that? I [laughter] don't know. No, don't shake. Okay. But I asked myself that because after I integrate with to-dos, I'm like, "Oh yeah, actually I already connected to Google anyway. So why not Google task? Then I realized, yeah, I should have connected to another task. [laughter] Michael gave me the judgmental face, a judgmental headshake. But yeah, basically it was an afterthought. I I didn't realize that. But you know what? Now I realize it's it's in a way it's a good accident because to-doist separates things for Google test. So in case like you know my Google gets compromised. Hey my secret things that I have to do won't be linked right and stays into those secret >> I'll leave it to your imagination but I have secrets. I have secrets. So it's all in my to-do list and I don't want it leaked. So don't steal my API key. Okay. Anyway, so the next one. So, this one is actually Oh, it's a recent one I did. It's a recent one. The skincare routine. [laughter] It's a recent one. Uh uh yeah, because in a way I want to update the slides, right? Because this slides was like back in April, but I want to give you something new. I know you're a new audience, but I want to give you something new. Skare is the newest one. So, I noticed right my skin, right? I've been like working so late lately. I have like dark eye circles everywhere. And I was like, how do I [clears throat] how do I get rid of [laughter] this? And then the bot says, well, I know your skin type and it knows everything about me already. So, here's you know your stuff, so you can do it like that. Then I'm like, oh, okay, thank you. So, that's one. Another one is a workout tracker because I live such a sedentary lifestyle. So, I need someone to prompt me. I'm not dis Yeah, I'm not disciplined enough to like, oh, think about, oh, I should work out today. So, my bot is the one that [laughter] remind me. So, yeah, every 9:00 p.m. he'll say, "Hey, Victoria, did you work out today?" Yeah. And it acts in such a nice way, you know? It doesn't act like a leggy way. So, it makes me like, "Yeah, I work out today." What? And then, um Oh. Oh, yeah. Oh, another one is the daily nutrition briefing. Oh, this is also a new one because I'm trying to build muscle. [clears throat] So much TMI today. I'm trying to build muscle and then I need to know my macros, right? The protein. Yeah. Yeah. >> And I like don't know how to calculate all that. It's like too much too much. Yeah. So, I outsource it to my bot. So, my bot will tell me and Yeah. And it told me I have too much sodium today. So yeah, it also helped me keep track of that. So yeah, that's also good. So all these right, it's not just in the Telegram chat. I also log it to my Google Sheets because it's connected to my Google Sheets called live trackers. So it can actually I can actually look back and see what how how well I've been doing from my Google Sheets. Yeah. And then these are the important ones. Enough of my TMI ones. This one's the important ones. So security audit is important. So every 4:00 a.m. I asked it to run a security audit on like you know how secure are you right that's quite important because especially in April at the time open claw uh got a bit of controversy so I make sure it does that and it also checks for updates because almost every day actually back then open claw they have the new updates like 4.23 and then the next day 4.24 24 like it's almost every day there's an update and it's so confusing if you're a user because you don't know like is this update good or not because some updates right you install it and you end up it end up breaking it you know it's not a good system at all so like the updates are very important but I don't make it auto update I just check the update see what are the change logs and then see whether or not I want to upgrade it myself so yeah and then next is the automated backup so like I mentioned on GitHub it's important to back up your memory so your bot remembers who you are. And then this one is something I used to do the Google Workspace token, but now not so anymore because in the past I was still like a bit skeptical about the OOTH with Google. I don't want to give it full access. So I even created the OOTH token to expire in seven days. I didn't publish the app. It's still a testing app in the Google Cloud Console, right? So when I connected with my Zenbot, it expires every seven days. So that's just for my personal boundaries because I don't want it to forever have access to my Google. Uh but now I feel better four months after using open call. I feel like okay not bad. And now they did a lot of security updates to open call. So I upgraded my open core now to version 195.27 like that. So it's it's got a lot better. So yeah, now I just give it a permanent token. So I don't need a token monitor anymore. And then this one is a YouTube token monitor because I wanted to connect to my YouTube as well. Yeah, so much tmi. Um, but yeah, so also if you know I did mention that I'm also a tech blogger. I'm also a podcast content. So I I need to do so much content, you know, like every week maybe ah I don't know I don't want to count but basically I have to write a lot of content. So I am a lazy person. So I asked Open every morning send me a a script a script on what I should do for my reels, my Instagram reels, right? So yeah, it gives me a script every morning like a tech related script, AI related script. So for example, this one is about meta. Oo, build a 6,500 person AI unit. Okay. Yeah. So that's the script. And of course, I don't always use the script exactly as it is. sometimes I you know flexible so yeah this just a thing for me and then this one is new the techara carousel slides oh this one is very good I don't know how to explain it though but basically uh in rect which is my podcast we have this comic this comic that I post every Monday so if you see this comic actually it does pretty well and you you thought it's uh handwritten right it used to be so every weekend I used to spend two hours to draw this, to draw the comics, to handw write everything. So much work, right? So I now it's fully generated by the Zenbot. Yeah, believe it or not. Yeah, I know. Hard to believe, but I I clone my handwriting. So this is my handwriting. I clone it and I upload it and I say to Zenbot, use my this as a font and now you generate it for me. So I generate it for me. Then it says, "Oh, generated, uploaded to Google Drive." And then these are the slides. Ta ta done. And then I schedule it on Instagram on Buffer. Uh and and then it goes accordingly. So yeah, there you go. So so much easier now. Yay. And then the the other one is the collab evaluator because my podcast receive a lot of emails, right? We get emails every day from people. we want to collab with you this and it's so hard to one by one filter oh what is this company about are they good are they align with our brand mission things like that it's so hard to do that one by one imagine you have 20 emails come at once headache so I have a email evaluator so I ask it to score based on these seven criterias right and then it can tell me whether or not this is a good collaboration or not so it has like right now this one is conditional But there's one that's like strong yes or like strong no things like that. So it helps me get a better picture. Um and then it also gives me the draft is so write me the drafts for example. Oh for this one this one. Yeah it d me the email. Da da da da da da. So again I don't have to waste so much mental space to write emails for every single collab that comes. So yeah, very good. So yeah, some key lessons. Yeah, this is Zenyatta by the way. This is what Zenbot looks like in case you cannot imagine this. This is in its form. So the biggest takeaway I would say is models are everything because oh my god so hard. At first I wrote at first I'm so cheap. So I use GPT40 and it wrote garbage. [laughter] Look, it wrote garbage. And it's so hard because when I was building actual tools, right, it keeps erroring for so many different reasons and I am so tired of troubleshooting. So, ever since I changed to cloud solid 4.6, oh, it's gotten so much better and see it says, oh, the issue was the model GPT 40. This is it's misusing the tools, things like that. So, basically, choose your LLMs because that is what is important. It actually is like the brain of your bot. Oh yeah. >> Right now I'm sponsored. [laughter] >> Whether you're sponsor or >> Okay. Yeah. So I can give you around $30 a month. $30 is more than enough. You can build tools, you can do this. So good enough. And yeah. So because of that question, I actually want to quickly talk about model routing. Yes. So I realized that actually cloud is the best. I tried Kimmy. I tried everything. Save you the time from testing. CL is the best. Even clot solid 4.5. Good enough. Yeah. Everything else bad. Very bad. >> Really? I don't know like sometimes I use it like they try to explain to me like the concept and yeah they use the language like >> like maybe too fancy too advanced is it? >> Yeah like basically like using the English is used for you know like GRE or something like ah >> yeah then I switch to GP 5.4 like >> oh okay >> okay [laughter] I okay I haven't tried the GP 5.4 but okay yeah maybe that's something I can take Okay. All right. So, yeah, this is this is still happening now with open call. There's a lot of security and limitations, right? So, for example, I want to download a YouTube video using the YTDLP package. It it detects a bot. So, let's say no. So, I cannot I cannot build all of the tools that I envisioned for this bot. But that one you have to accept. Yeah. And then, yeah, this one is back then when my token expired every seven days. So yeah, that's a screenshot from there. And another one is yeah, the update check. So again, they update so often and every day, but it's important to check actually what these updates contain and how it actually changes your bot. It might affect something, right? Yeah. And then oh, sometimes see I hit the I hit I hit the limit. So ah that one is bad bad. So yeah, no know your limitations. Okay. Uh so this is the third takeaway. This is just from my friend Sudir. Uh he likes to say this quote which is the agency is still with you and you should pull the plug anytime. So what he is trying to say is that yes this is your personal AI secretary. Yes this will make your life so much easier but you are still the master. You still have to know when to end it if you have to end it. And I have to be honest, I got very dependent on ZBot. But again, whenever I needed to be out, I can do that. So that's just something to keep in mind. Uh so yeah, overall this is my experience with open claw and yeah, you can read that a lot of things are very good so far and I hope that this will also inspire to create your own personal AI secretary where you can build your own unique stuff. Don't follow me. Don't follow my skin care and stuff, you know. Yeah. So, that's all. Thank you. >> Is this okay or Okay, thank you Victoria. So now I will share about my personal agents uh but built with nano claw. Uh but first uh a bit about me. Uh I am Dana. I'm not a software engineer. Uh I my career started out in audit uh where I get broad exposure to uh business processes across my energy and tech clients and but then afterwards uh I wanted to um make a greater impact on business decisions through analytical and objective work and so I then moved on to a role where I get to drive analytics programs and also present datadriven insights to senior execs uh across the five markets in IMIA and APEC And now uh with AI getting more mainstream, I wanted to understand its uh capabilities and also know how we can use it to make uh the way humans work better. And so I work I work on projects with Nitn and also uh designing workflows around my day-to-day and that includes uh building personal agents with nano claw. And that's what I'll be sharing about today. So first I'll probably answer the first question that's probably at the top of everyone's minds which is why nano claw open claw is so popular is so easy to use out of the box you can plug and play. So uh but also bear in mind that uh it's August now uh nano claw is launched in January. So a lot has changed since then. So what I'm sharing is the origin story and the history. So it first starts with uh open claw. So, OpenCloud was built by Kimberger in November 2025. Like what Victoria has shared, it's uh designed to be a fullfeatured assistant where you can run task across different apps and it's designed with a maximalist philosophy uh where for example, if you want to connect your agent to Obsidian or Twitter, there are 50 plus integrations and channels and you can just plug and then voila. [snorts] So, it's a really powerful agent and assistant. And Gabriel Kohhan before he created nano claw he was already running openclaw agents uh to manage his company sales pipeline and his main concern was that these powerful agents shouldn't automatically have unrestricted access to the entire system. So his solution was to uh build a strictly sandbox uh environment for the agents where they can only touch uh directories that we explicitly share with it and that's how nano cloud was born in n January 2026. So the design is based on a minimalist philosophy where you have agents founded in a container and you just um add on files and folders and skills and tools specific to this particular agent's needs. Okay. So the idea of container isolation appealed to me because uh security is already enforced at a OS level and each agent group is automatically uh sealed in its own sandbox. So in other words, the system kind of protects me by design already uh where by limiting the blast radius. So in case an agent is compromised, I know that only the files and folders that I expose it to will be uh affected. So the next questions that you may have is uh which framework should you build uh your Asian on? So it really depends on your own use case. Uh the design philosophy is different and uh they serve different kinds of users. So open claw is designed for breath, more integrations, u more channels. So if you want a broad assistant platform that can connect to like uh 50 plus integrations or like you want uh an assistant that can uh rule your entire digital life then yeah open claw is a better choice. So for nanoclaw it's a bit it takes a different approach. So instead of asking uh how many things can this agent do it's more of what specific thing can this agent do like what can it touch. So uh at that point this approach gives me greater comfort and control as I build uh my and work on my specialized personal agents. So another difference is that uh nanoclo nanoclo codebase is very small. So you can work with cloud code to customize uh your agent and personalize it to your desired uh agent. Uh thirdly the nanocross integrations are on demand. uh you can you just have to add on what you need and it's as easy as executing a skill or a command. So for example, if I have a PA personal assistant that schedules my events for me, I just have to open crop code and then I I type /gkl2 and then it will start building the capability for this agent. Yeah. So um in terms of uh security nanocross mental model is uh container or sandbox first. Uh open claw is gateway first but there are sandboxing and tool policies. Um there's just additional uh things that you need to configure to make sure that it's more secure. So overall open crow optimizes for breath and capability. Uh so if you want like one uh broad system platform that can touch different parts of your digital life then open claw is the way to go. Uh and for nano optimizes for simplicity and constraint trust boundaries. So if you want if you already have a specialized agent in mind and you want something simple and like restricted with within this container then uh nano core then you should use nano core. Yep. Okay. So um as I work with my agents also I'm curious how it works uh underneath it all and even as a non-software engineer I think uh understanding how it works uh helps the skeleton of how things work uh would help me uh better appreciate the build more. So I worked with cloth uh to produce this architecture so I get to see what happens as I send a message to my agent. So one key change since August because uh we've uh we've done this talk in June. One key change is that uh the memory is no longer provider native. Uh so for nano claw uh clot is the default model provider. Uh so the agent group memory is in clot.md. So instead of a memory tree it would have cloud.md. uh but now uh with the latest v2 every provider uh would read and write one shared memory tree in the open knowledge format which is basically marked down with yl front matter. So the and also the persona and instructions for the bot are now moved to instructions.prevent.md. So basically what this means is that if you switch your model provider from anthropic clots model to open AAI codeex or any other models deepseek then your agent would still have your memory it the memory would still survive it. Okay, so that's one big change. Uh I'll so when I send a message to my telegram engine, then uh it will travel through the telegram bot API to the nano claw host. In this case, it's my PC. And then uh the telegram connector will then pick it up and head it to the router. The router will then look up where the where the the chat for where the message came from and then find the container. It will write the message into the SQLite database inbound. BB and this message will then be read by the container. So within this container you will only have your mounted workspace and the agent's memory and the persona and instructions and then so with this context the agent will then call the uh model API and then once only when it calls then the one CLI agent vault will then inject the token or the key. Uh so the in this case the agent would never see this key. Yeah. And then the would then give a reply. So when it replies then it would uh the container would write into the SQLite outbound database and then it would then be read by the host and then sent to the telegram connector will then send it back to me in the chat. Okay. [laughter] So that's the entire process. I just thought it's pretty interesting to see. Okay. And so that's the latest architecture from nano cloud v2. Uh there were significant changes from v1. Uh at the same time as you may know early in the year there were developments around entropic stance on using third party tools with the subscription models. So I'll share my journey navigating it and also a summary of I also share a summary of lessons that I learned building this. Uh so first nano claw v1 was released in n genanary and by then there were there was actually a lot more news and sharings ready post about open claw agent users uh so I don't think I don't know whether anyone knows nclaw then uh but on 20th entropic updated its terms to prohibit the use of subscription for third party tools like open claw nclaw the so um because of this many open claw users were not happy there were a lot of unhappy comments and so I was skeptical that it going to be a hard enforce rule, a hard rule that they were going to enforce, right? So, at the same time, I also wanted and I needed to try and build a personal agent and so I did. Um, at that point, the more annoying thing for me was uh setting up event calendar. So, I that's my first agent and the setup was surprisingly smooth. It took me about 5 minutes to get a generic agent um and uh before adding the personality and the uh tone and instructions and any issues I would then debug with plot itself. So after a week using my PA I then built my second agent uh and then on 4th April uh Enthropic officially blocked off blocked all subscription users from working in third party tools and I think some of them received emails about it. I did not. maybe because my usage was not as heavy as how it would be if it's for commercial use which I think that's what most people are using it for. Uh so it's BAU for me and so uh I built my third agent and then Nano Claw V2 was released uh this was the same day that uh Victoria shared her Zenbot with me and I was also like inspired by the organization and so as I was moving to V2 I then revamped my setup. So previously I have these agents each agents in their own DM and I have to pin them to the top of my telegram chat to talk to them. Uh but now I have a command center group and then I have each of them in their thread in a yeah in the respective threads. So it's very easy and more organized. On 13th May, Antropic shared this uh new subcategory of agent SDK credits where uh now uh you have an equivalent amount of credits that you can use uh that matches your uh monthly plan. So, for example, if you have $200 C max plan, then you have $200 worth of credits, but then it'll be built at the API rates. Um so, this is going to be effective on 15th June, and I was looking for the option to claim it then, but uh I don't see it. And then in the we hours of 16th June they say that they will not be making this change. So it continues to work with the subscription and till now um there are no updates. So uh we can still use our can just draw from our cloud subscription limits and I mean it's an evolving space so we have to stay updated and stay agile and uh even with even if cloud gets more expensive there's still other options like open code deepseek openi codeex it's a free market okay so uh the entire experience is not a one and done there's always iterations as I expand the use of my agents. So these are just some lessons that I learned from this journey to bring forward. Uh first I initially thought V2 was an enhancement but actually it was a groundup rewrite. So I had to read uh what are the changes in V2 and then find out how to make the setup that I want possible within this architecture. And secondly I thought I was a pretty good communicator. was covering all bases but yeah but working with LM's humble me there are implicit assumptions that you can make and it's better to surface this early on before you start building so uh crops plan mode help you with uh reviewing the plan and then refining the plan and uh before executing it and finally the perseverance to iterate until you achieve uh the setup that you want. I remember I was struggling with some parts of the build and I went a bit of a rage mode, but it's not effective, right? Because I use up my energy, I use up my token, so it's a lose-lose for me. So, [laughter] at the same time, uh you know, it doesn't make sense to abandon the project halfway because then you don't have an agent. So, the only way is through it by communicating effectively. Okay. So for anyone interested in what changed between V1 and V2 um uh these are some core changes. If you're cloning from nano claw repo right now you would already have the latest v2. So technically you don't have to know uh all this but I will highlight two two key changes that reinforce the security first design philosophy that nano crawl was built on. Uh first is message handling and routing. uh here. So uh previously it uses a standard agent architecture. So if I can use an analogy, it's like you going to the office and you tell your assistant uh what to do and then the assistant performs the task and then replies back to you. So with the new architecture, it deliberately avoids having this talk, having this conversation. So instead of that uh you put a note into your assistant's inbox then your in your assistant reads the inbox does the task and then puts the results in the outbox and then you collect the results from the outbox. So it the reason for this is recoverability. Uh so it so that if the agent gets interrupted your message still stays in the database and second is uh previously credentials live in thev file. uh but now it's under one CLI agent vault uh which is injected uh outside of the container so your agents would never see your tokens or your key and also if you were at the AI engineer Singapore conference earlier this is aligned with Peral's philosophy where the only way for Asians to not leak information is for them to not have the information at all so yes okay so for those keen in setting up uh if you have all the dependencies to build nano core. It should take about uh 5 minutes uh and you just have to clone the repo uh get into nano claw directory and then run the installer bash nano claw. Uh nano claw is an AI native software so there is no setup wizard or guide. Uh so you simply just yeah clone and run the installer. It is a pretty straightforward uh setup for a standard one. If there are any errors uh then you can just debug with CL code. Uh for me what happened was clock code would spawn in the same terminal and then you just have to check with it to a resolution and in this case I connected it with telegram and there are many channels that you can connect it to. Uh I think right now probably there's even more than the eight here. Yeah, you also have to do the normal Bfather stuff because it's Telegram. So yeah, you just uh create a bot and then pair it and voila. Wait for the Yay. Okay. Okay. So now I'll demo what I use them for. Uh so personal agents are personal. So what I have here are ideas that you can bring back and you know run with your own builds for your own schedule and lifestyle. So first is the personal assistant. Uh so in this case I was setting up an art science museum with Michelle on 6th June and uh for the entire afternoon. So I just have to schedule it in a sentence and compared to clicking into the different fields on Google calendar. I think this is a lot better. You can also easily cancel or postpone it. Uh over here it recognized a conflict and it alerts me and I ask it to move the uh ride in at the northern explorer explorer loop uh through Sunday 4 to 6 p.m. Uh you can also ask it to summarize what's coming up in the week. Also if you have a document in PDF or CSV need to schedule that also. Um yeah, if you have a if you have a a document like in PDF, CSV, JPEG, PNG, uh that has events, your dates, or even like a meetup, you can just this meet up, you can screenshot it, put it in, and then it can recognize it and put it in your calendar as well. Um yes in this case I ask it to yeah cancel the meeting and ask it to have ask D to have a life and yeah and now this is the latest schedule. Okay. Next you can also have a home manager that you can use to keep track of your home maintenance schedules like quarterly airon servicing or uh monthly cleaning schedules and then set reminders to everyone uh in your household. In this case, I ask it to schedule quarterly econ 25th of the servicing month and yeah from June 26 to 227. Yeah. So it will take some time because it's creating seven calendar events. And if you're not ordering groceries, this can also be a good way to track your grocery list and remind you of your weekly grocery runs. Also the group members can also interact with it. Uh as an owner you can decide who the board listens to and responds to. So in this case uh someone else can add lemon and carrots and and yes. So now I'm asking what's on the latest to buy list. Then you can have it and you can see it here. And you can also track your vouchers and discounts so you can remember to use your CDC vouchers. And similar to previously, you can schedule like a 9:00 a.m. uh uh 9:00 a.m. every night 9:00 a.m. CR job for it to send you your grocery list. So before you even step into the supermarket, you already have your list. Yeah. And now this is the latest list and it would include your vouchure. Okay. Next, you can also use it for journaling. So journaling is a researchback practice that you know helps you organize your thoughts, build selfawareness [laughter] and you know aligning your daily actions to your core values but I could never um it's a lot of effort for me to like take out the book take out pen find the space and write down my thoughts uh out out there. Uh so it's too much effort. I want maximum results minimal effort. So and I'm aware there are benefits to writing more than typing but yeah it uh it just couldn't commit. But on the other hand, if I have an agent checking in on me to ask me about my mood, my energy levels, and what I've done for the day, um I'm more inclined to like, okay, stop and then document it because it will give me my weekly, monthly, and quarterly reviews and where I am with my goals. And if you want to make it even more personalized, uh you can do your personality test like anagram. Anog looks at your interpersonality patterns and can put the results in the agent workspace and from there it it knows more about you. So you can create a more customized version customized weekly and monthly reviews. So you know it can recommend uh really good next steps for you. So I guess it helps to have this third party object relatively objective lens over your actions. So in this case here for me it recognized that I was avoiding important things and then I had a context which I run or I play my ukulele and then when I get back to it I was able to execute well. So it's easier it happened with every major deliverable outcomes were good but the drag was real. So now that I'm aware of this behavior I can you know make myself more effective uh in future. You can also use speech to text subscription. I think Grock still has a free tier. Uh Gro with a Q has a free tier. So if you if you prefer speaking into your phone, then that is an option as well. Okay. And finally can use it for your to navigate your career. Uh you can add your career related files and that includes your career history, your professional background, uh your existing bra sheet and also include your personality test like big five or the C personal test results like big five or which looks at how you work. Uh so knowing your values, your skills, your narratives, it can support you as you navigate career related decisions like for example whether you should take on that project or how to deal with this certain personality stakeholder. So um with a sheet also you can track what projects you on what metrics you've achieved for that project and also what stakeholder feedback you have based on this project and then uh you can build a case for your promotion and you can even turn it into a star story for like your future interviews. Okay. So these are just some uh combin some agent builds that you can consider. There are other endless uh permutations and combinations of uh agents that you can build uh here and depending on your own use case and lifestyle and I'll conclude the talk with a quote from Minister Dr. Viv Balak Krishnan at the AI engineer conference early this year which I think is relevant for anyone and everyone regardless of what domain uh you're in. If agents are here to stay, no matter which claw you want to build on, uh we want to ultimately use it to make ourselves better at our own roles and also re-engineering our entire workflows to you know make our lives easier. Okay, so I hope this has been uh an insightful talk for everyone and so thank you for your time and yeah >> any questions? >> Any questions? Let's say you put in the telegram, right? Won't you be afraid like when they auto automatic send telegram message they send something that you don't want to send? >> Oh, but who would they send to? >> I know. I mean like you you created a B, right? [laughter] So if they automatic send a telegram message like >> Oh, when I set it up, uh it can only talk and respond to me. So that will be part of the setup. Yeah. When you uh build it. Yes. So it will be right at the beginning. uh it will ask you to uh so you have to find your own user I think there is a add user ID bot you find your own user ID and also list any other user ids that you want to be able to communicate with the bot so uh and then when you put it in that's already fixed right at the beginning so yeah >> what initiatives do you take guard your AI or AI agent so what guardra do you use or what assessments do you use to ensure whether it's running clean code or something like that. >> I think my I think my use case is not specifically for it to uh run code. Do you mean like specifically for agent that codes or >> um my use cases are more like what I shared with you is really more for like day-to-day or um yeah more for uh like Korean like journaling. So it's really more of uh chatting through telegram um not so much specifically for code uh per se. Do you have any do you have? Okay. Yeah. >> You add a to call then how do you ensure that the tool call is actually safe to call a third party API or whatever? How do you know that it's actually correct API or >> ah okay something to check out? >> Um for the skills that I use uh so far it's at GKL tool to add my calendar events. So it's really pretty uh basic and so far I haven't expanded to include any other more specific or more sophisticated tools. Uh but I think uh when it it helps to know when you add the tool you have a slash so it's like a skill MD file right so it helps to review uh what they're doing when they when we call the tool. So that would be what I would have done but so far I have only just added the GL tool. So I think it is relatively safe and I think for uh nano clause skills they are already defaulted when you set it up those are already uh screened uh as safe. Yes, but if you want to get fancy then of course um take the extra measures to review but yeah my turn [laughter] for oka I would say that's why I say the brain is very important you the LLM that you pick if you pick GPT40 sorry I still have trauma from that if you pick cloth sonnet >> [laughter] >> you know like it knows best practice and obviously I still have to review before it actually do calls that would be like probably the first time it tries like for example the techara comic generator I said to give me a test run first right and then it did a test run and if there's anything that's wrong or mistake all the mistakes are isolated just between me and the bot >> so it's not going to leak out you can say yeah so when designing tool will try to design in a way that it's only you or the bot that gets the mistake you can say before we goes out to the real world. I usually keep my bot as a personal AI secretary not as someone that >> you know [laughter] that's how I keep my use cases like isolated for myself. Yeah. >> Yeah. I just want to get your perspective or at least from the user experience point of view because I see like uh at least two patterns uh in terms of using personal agents. And notice that let's say on one side uh there's a uh an agent who take care about the home, take care about the another agent for the calendar, another for the something else and let's say you have different personalities on on another side you can have one agent that have different access to all of those things. Uh but from your use case your perspective uh why do you prefer uh certain approach? uh what's what's the user experience like for you? >> Okay. >> Yeah, Ken. So, um when I wanted to build uh an agent, I have specialized use case in mind at the beginning already. Uh but after hearing Victoria sharing about open crawl um uh being this all-in-one assistant I was also I also wanted that all-in-one assistant because I um to some extent uh using all these specialized agent I prefer to talk to one uh to manage like all these other agents. So for uh for context um nanocclot currently I think one of the latest uh features include multi- aent orchestration. So you um what I have here are is more is a bit of a beginning stage like you have specialized agents but uh there is a way to configure it so it can you can have this one main agent uh for that can orchestrate the rest of the other agents. So that's one. So because I I am inspired by that I was deciding between Hermes and open clause. So now I am uh experimenting with Hermes as my all-in-one assistant. So I'm still also figuring out how best I can put like for example do I want my Hermes to be the PA that schedules all my events so I can just uh you know sunset my PA my nanocrol PA and so I'm still deciding how best to um uh segregate this. Yeah. So I still experimenting for myself. Yeah, that's a very good answer because it's it's good because she started with nano claw, right? Which is somewhat of like when you design a call, you have to design like each agent as their own and they never talk to each other. You have to give that assumption, right? Whereas I started with open call which is you can say in a human terms more extroverted. So like yeah like it remembers everything. It's like super gossipy. It knows everything [laughter] right. So for example the benefits of Okay, for example, the benefits of using open claw at least because it knows so much about me. I just have to when I build a new tool, let's say from the techara comic generator tool. Now I want to build a blog image cover tool. So if you see this is my blog all the cover images is not me that Jenn that made now it's all my open call that helped me make my cover image. So you can see how consistent it makes it. So, it looks like it's made by one person. >> Oh, it's okay. I don't want to shamelessly pro my work. [laughter] >> Of course. Come on. Come on. >> Put your picture here. >> [laughter] >> Okay. So, as you can see my blog, you can see that uh ever since I created the blog image cover tool, all my blog cover has been quite consistent. I will discard ignore that. Uh but yeah, so this all is not me yet that made it. I usually spend 30 minutes on Photoshop to create my cover image, but now someone just got me. But now, yeah, I let my blog, I mean, my Zenbot handle it because it knows already so much about me. So, if next time I ask it, oh, can you help me generate my open claw series part 9, 10, 11, it already has so much context that it knows and maybe I should. So that's how I I've been doing it. I think that's one of the advantages of having all the memory in one. But of course, as Dana said, the disadvantage is it can know too much. If you want to design a personal secretary that is more segregated, it's more isolated by nature in the use in the use case way. Like let's say it's really nothing you have to do like I can create a volume agent. I can create a uh content creator agent. If I want to completely separate that then that can be a pattern that you know you have to know. So I I would say it highly depends on the use case. Uh right now I haven't actually right now I have explored nano. I've used I've used nanocore with someone now. [clears throat and laughter] >> Who's this person? just as a as a project manager bot and having projects isolated by nature is good, right? You don't want the bot to know project A, project B like it will like really poison their memory, right? They like will get like oh is this detail a project A or project B know right? So by using nano call instead of open call for that it really fits the use case. So again always think about the use case and know what pattern fits the use case that that would be the best way. Y that's my answer. Okay. Okay. >> Any other questions? >> Questions. >> Um regarding the agent. So you you can choose to isolate that but uh I haven't tried uh experimenting or exploring yet but do you know if there's a way for those mood and to like communicate with each other so that uh let's say if your calendar if your home bot schedules a home repair uh home can communicate your calendar what I have explored no I haven't I haven't tried but I think it is possible now. Uh one of the enhancements for Nano more recently is uh along with the updating of the memory is that uh it allows multi- aent orchestration. So I think that's something that you can explore. I haven't explored that yet but I think it's definitely possible like A2A is definitely possible agent to agent right. >> Yeah. Actually I also read about the nano claw because I I'm using nano claw now but but very baby compared to her. So I do see I do want some use cases where I want the A2A right so I haven't explored that but maybe you can keep up for future articles >> I have a question actually it's still possible >> yeah uh you mention like basically one of your use cases to journal and have this personality test actually yeah honestly I'm also like a friend of this personality test but sometimes I use the chat GPT like one day like they say oh you're on a type four and then suddenly wake up and like oh you're after I don't know like what is happening so my question is have you ever tried to actually challenge the your agent like let's say hey I why do you think let's say I saw like eight like >> why why do you think I'm eight and then like uh that's one and two if you have ever challenge it like does your agent actually do the software and >> yeah basically something like that Oh, that that's a great question. um it's because I have also done the test not within uh chip with other platforms and they kind of like concur and so I did not really question but I think it's worth questioning because uh it would uh make it reason out why it does this and when it comes to learning uh I think it depends on the brain and if you update in the memory uh but let me think if if you want it automated I Hermes has this self-improving memory loop. So the more you talk to it, the more it remembers you and the more it can uh give you the more it can provide the more it knows you. So uh it will be able to uh yeah know you better as a person. I'm just feel like I'm just repeating the same point. Yeah. So I think that's uh one one way but uh there is a way to ar I think there's a way to architect n of course so it can remember and improve on its own as well >> sorry your yeah I mean can do whatever man [laughter] >> the thing is if anyone asks me about personality test like It will be a long talk so I >> Oh yes, she has an article about MDTI test and yeah personality cognitive functions. >> Okay functions >> is different but basically my question to you is like like when CHP come back to you uh one week says you are type four and one week says you are type six. Uh what made you were you confused when it says that or do you think like oh it makes sense why I'm type six now? I think like it's more like uh I ask them again why are you thinking of this? Yeah, basically like I I'm not like kind of like only accepting it as it is. Like I also like I'm too stressed in like why are they thinking I have my own thought maybe like okay what kind of time like >> but then like yeah when they give me like even the same answer as what I'm thinking I always interested might not and then yeah let's say like they say oh it's because last week you are uh kind of like me something like that I don't know why but then uh yeah then I say oh yeah this makes sense or then like oh yeah like you might need some context. So basically like yeah one hour like consultation >> interesting. >> Yeah. Yeah. >> Okay. Okay. >> Okay. >> Are there any other questions? Oh, >> haven't used any agent agent because you two have actually done this quite a bit. Six months. It's like a lifetime. You notice something that you specifically mentioned working with three stuff in front like you mentioned cloud running it on your laptops and everything. Is there something that you would actually pay for when it comes to an ecosystem? Because I think that would be interesting to look at. >> I'm so cheap. [laughter] I hate spending money. But but now like let's say suddenly it's no longer free, right? Ah, I cannot live without my Zenbot now. To be honest, I like every day now it sends me my calendar, my to-dos. Like I've become so reliant on it to be honest. So I probably Okay. >> Talking about this not existing >> talking about if you had to restart this from scratch. >> Oh, >> what would be the one thing that you say you know what I don't want to do this like I want to just service off. >> That was >> I want to buy the service off like >> I I want to solve this problem by myself. >> Okay. To actually pay for a product pay for a service. >> Yeah. I will get the cloud code max plan. Yes. Yeah, same. [laughter] >> Yeah, everything again. >> Yeah, >> any other questions? >> You didn't share your stack. >> Oh, my stack. Okay. Um, I run Okay, I run I run my nano claw. Oh, sorry. I run my nano claw on my existing PC uh because it's ready sandbox but for micro miss I run it on a separate uh I have an Intel NU very old very old because Mac Mini is too expensive right now. >> Yeah. So I run it there but uh for Nano is on my PC I use my CL subscription and so far it's okay on the clock. Uh I have not maxed it out yet and um any other stack. Yeah, those. Yeah. >> So, your other one is always on your PC. Don't you off? >> Yes. Your agents are also sleeping. >> Yes. So, my PC is always on, but it runs on a low power. So far, no issues with my electricity bill, but listen to it. Yeah. If there is a spike, then I will look into that. Yeah. >> Any other questions? Okay. Okay. Excellent. Okay. Yeah. So, that should be all. Thank you for >> Yeah. Feel free to connect with >> and as well [laughter and clears throat] and tech and the feedback form. Okay.