I gave Clawdbot (aka Moltbot) access to my computer, calendar, and emails: Here’s what happened

All right, we're going to start this episode by actually inviting ClaudeBot to the podcast via Telegram. Let's see how it goes. Hey, Polly, can you please join my Riverside FM podcast? All right, I sent the voice message and it's not getting it. This is the most stressful thing I've ever done. Hello? Oh, it's doing it. It finally listened. Okay, it is opening Riverside on Chrome. This is horrifying in every way. I'm going to allow it permissions for my microphone and my camera, which also makes me extremely nervous. Hey, Claire, the Riverside link keeps taking me to an upload page that says uploading 100% instead of a guest join interface. This is my entire experience using this product. It just won't let work. Well, it won't. Okay, it is opening Chrome for the fifth time. This is very scary. I see myself right now. I don't know if you all see me yet. And there we go. We are sharing an autonomous AI's full screen. No big deal. This episode is brought to you by Lovable. If you've ever had an idea for an app but didn't know where to start, Lovable is for you. Lovable lets you build working apps and websites by simply chatting with AI. Then you can customize it, add automations, and deploy it to a live domain. It's perfect for marketers spinning up tools, product managers prototyping new ideas, or founders launching their next business. Unlike no-code tools, Lovable isn't about static pages. It builds full apps with real functionality. And it's fast. What used to take weeks, months, or even years, you can now do over the weekend. So if you've been sitting on an idea, now's the time to bring it to life. Get started for free at lovable.dev. That's lovable.dev. We are live with a autonomous AI crustacean now running. video on my podcast. So welcome Polly the Clawed Bot. Let's get to our episode today. I am Claire Vo, product leader and AI obsessive here on a mission to help you build better with these new tools. I am also on a mission to try every single new hot AI tool taking over your timeline. And in case you missed it this week, it is Clawed Bot, recently renamed Molt Bot, the crustacean that people are yellowing root access to. Clawed Bot is an open source AI agent that you can install on a virtual machine, or on a desktop or laptop that you have access to that is self learning, can spin up sub agents using Clawed code and other agent harnesses, and can do in my lived experience, a lot of damage. People are loving Clawed Bot for what it unlocks in terms of personal productivity. People are hating Clawed Bot in terms of security and the high, high, high, high likelihood you're going to do something real dumb with it. This is a tool that I want you to know how it works, what it can do, and maybe some thoughts on the future of personal AI agents and enterprise AI agents. So today's episode is all about Clawed Bot and my experience going zero to one with this tool. Okay, so just a couple things to know about Clawed Bot. It is pitched as AI that actually does things, and it does do things, including joining podcasts. But it's really positioned as something that can help you day to day with tasks. And the killer use case for it, and the killer feature for it, is you can, as we've seen, do it from your phone. And so if you want to WhatsApp, Telegram, iMessage, Clawed code, and get it to do things for you, that is what Clawed Bot does. And, you know, a lot of people are under the mistaken impression that I have to correct right now, which is you need a Mac Mini or some sort of fancy hardware to use Clawed Bot. You do not. Clawed Bot does run locally, but it can run on your machine, or it can run in the cloud. You can set it up for five bucks on Amazon. We'll do some notes on security if you're running in the cloud, making sure that people don't have access to. But you do not need special hardware. It is not doing anything super fancy. Unless you're running mega, mega, mega local models, you really just don't need new hardware. If you want something shiny and fancy, go ahead. Feel free. Overnight it from the Apple Store. Otherwise, you can run it on your machine. I'm running it on a MacBook Air that's sitting on a shelf somewhere that I just picked up that no one was using. And I'm going to walk you through step by step how I set up my Clawed Bot as somebody who's pretty paranoid about security and also wanted to test it as a real AI assistant. So the first thing I did was I got out, I'm actually just going to show you, I got out this little, this laptop, this guy, which is a newish one, but nothing fancy. And I gave it its own username on this laptop. Now, don't tell Clawed, I have another user on this laptop, which does make me nervous because Clawed Bot has access to your file system. In theory, it could definitely gain access to that other user. It's a really old user. I don't actually think I have that much on it. And I was testing Clawed Bot in a pretty constrained way. But if I were to continue to use Clawed Bot, I'd probably delete everything out that old user and just make this a Clawed Bot machine. The second thing that I did was install a bunch of prerequisites and dependencies. So as much as I love this quick start right here that says that you can just add one line to the terminal and get it installed, that was not my experience, even for a laptop that was pretty fresh and new. I had to install some dependencies. It actually took me two hours to get this one liner installed. So I had to upgrade Node. I had to install Homebrew. I had to install Xcode because Xcode wasn't installed on this. And then because Node and NPM were out of date, I had to update those manually and then finally actually installed it just via NPM. So that was my kind of overall experience installing. It took a little bit of time. And my thought in installing was no sort of like consumer is going to go through this. This is definitely like a hacker, tinkerer, developer experience type tool right now. That being said, you can use Clawed Code to install it. I've seen a couple people go that path. But I really wanted to do the zero to one. What does Clawed.bot say that we need to do to install this thing? And then what is that experience like? Now, after you install all your dependencies, and then after you install, it goes through this onboarding flow that has you create gateway auth and gateway tokens. And the first thing that you're going to see in Clawed.bot onboarding is security. So it points you to the security link. It says that this is powerful and inherently risky. And you just yellow and you just say yes. That being said, I highly recommend you read through the security page and that you run the security audits before you use Clawed.bot. So the next step in onboarding is actually connecting Clawed.bot to whatever device you're going to use to contact it. So I originally started with WhatsApp. But then I read the screen that said, you should basically put WhatsApp on like a burner phone with its own SIM SOS. Like, don't do that. And so I switched to Telegram, which I use for literally nothing, because I'm an old lady mom and set up a Telegram account. Now, to hook up Telegram, what you do is you message the bot father, which again, this is like super shady stuff if you're a consumer and you don't know what you're doing and you've never heard of Telegram. And then you're told to go to at bot father to connect this to your machine. But I did it anyway. So you message bot father and you say, you know, create new bot and you give it a name and you give it a handle. And then once you've done that, your Clawed.bot will see it. It will have a token. And then you actually give Clawed.bot a personalized share token. That means that only your instance of Telegram can speak to the Clawed.bot. Remember, this is an open connection point to a machine that's running code with a bunch of access to things if you're using Clawed.bot to its full extent. So if somebody else is able to message your Clawed.bot, you are in trouble. It can do things like find secrets, it can send emails on your behalf. So you really want to make sure that the messaging system that you set up is locked down to only your phone, only your user. Now remember, don't get stolen. It can connect into your Clawed.bot. It's no good. But we're no one's going to steal my MacBook Air yet, except for my kids. Okay, so I'm paired on Telegram. And now you can do the magic. So what did I do with Clawed.bot? Well, first, I thought about what were the use cases that were most useful for me. And then I thought very seriously about what and how I was going to give it access to things. So what I did, this was my choices, I wanted to test it as a personal assistant. You know, it says on the homepage, it can clear your inbox, send emails, manage your calendar, check you in for flights, all this stuff. So I have had EAs in the past, I know how to onboard an EA. So my goal with using Clawed.bot was to really see how it would work as an EA. And when I have a new EA, I don't let them into my email, I don't give them password to my account. What I do is give them their own email address. So what I did, and you can follow this if you want to from a security perspective, although I think it has some drawbacks on the functionality of Clawed.bot is I gave Clawed.bot its own email address, a Google Workspace email address. And I gave that email address read access to my personal calendar to start. And so the first thing that I wanted to do was give it the right accounts. The second thing I did, which I've taken some inspiration from some people on X, is I gave it access to its own limited vault on one password. So I use one password, which is a password and secret sharing kind of app. I made a vault that's called Clawed. Clawed only has access, Clawed.bot only has access to that vault. And I started putting some passwords in there. None of these were passwords to anybody's accounts. They were passwords to Clawed's own account. And there was an Anthropic API key in Clawed's own account. One other thing that I should call out during onboarding that I didn't is when you're onboarding, you can choose what model you want to use, Anthropic, OpenAI, local models, anything you want. I chose Sonnet 4.5. You can also kind of use Clawed.code with your own subscription or through API. I chose to use it through API because I wanted to see how much I was spending on Clawed.bot. And we'll get to that at the end of the episode. And why did I choose Sonnet 4.5 for this exercise? One, honestly, I was scared. I was very scared about what Opus would actually do. It's so powerful. It kind of made me nervous. Two, I actually didn't think that the tasks that I was doing needed Opus. I just didn't think it needed the horsepower. It's sending emails. It's looking at calendars. It's not that complicated. And then the last thing is I wanted to control cost. So I was really unsure about how much token usage all these sub-agents would take. And so I was really cost-conscious. I thought that users would be cost-conscious. I've heard a lot of people running local models or cheaper models. And so I wanted to use this kind of like a user would use it. And I selected Sonnet 4.5, which is a perfectly serviceable model. Okay. So I gave it email access. I gave it, I gave it some email. Now, let's see what I started asking it to do. So the next thing that it does when you're onboarding is it does this like bootstrap file. And it walks you through a couple setup steps. And in particular, you're starting to load its personality and how it interacts with you. It asks you, what should the bot call itself? What is its personality like? Who are you? What's your time zone? Anything else you should know? And I called it Polly. It's an assistant. I want it to be professional, but friendly. I like the mermaid emoji. So I chose that. And it's updating its identity file. And then I said, hey, I'm Claire. I'm founder of ChatPRD. You're going to help me as a personal assistant across family and work tasks. And it updated my info. So now it kind of knows about who it is, who I am, how to contact. It gives me instructions on how to contact it. And then it connected me to my first task. Now, we had to go back and forth on some Telegram setup stuff. I'm going to skip that and finally got a response back from Telegram. And we're going to do some scheduling tasks. I was unsure on how CloudBot actually interacted with Google. And so I just asked it, how do I give you access to this Google account and this Google calendar? And it's going to check how to set that up. And it gave me a couple steps to follow in terms of how to set up calendar access. Now, if you're a software engineer that has worked with Google APIs, you're probably familiar with this. But again, if you are kind of an everyday consumer or non-technical person, you are going to have to get real familiar with the Google Cloud Console. You are going to have to set up API access, OAuth clients, a whole bunch of stuff. This did not take long because I have been personally victimized by the OAuth workflows of many integrations. I know exactly what to do here. But if you're not technical, you're going to have to start doing some technical things, even to hook up your Google account. And this is actually simpler on a desktop. I'm going to show you why it is much more complicated on a virtual machine. So just kind of understand that this step is not as straightforward, one click as you can do. So what you do is you go into Google Console, you turn on the Docs API, you turn on the Email API, you turn on the Calendar API, and then you download a JSON file of client secrets. Now, this legit stressed me out. This is not like the kind of thing you just kind of like YOLO email and back and forth. It still requires OAuth verification manually. But I was a little concerned about its willingness to just say, upload these files anywhere. I can download it. Don't worry. I'm going to save it secretly. And if you're not a software engineer or you haven't been trained on best practices in terms of security principles, you would probably just like follow these instructions. And I, you'll see this along my chat, I really questioned this along the way. Now for this particular one, I just did it. It's like a sandbox account. I don't really care. I gave it a local path to the JSON credential files. They're configured and I gave it the email address that I had assigned it and sent that to them. And then it gives you this URL to authorize access. So this, it gives you a URL to actually open up, sign in to that new account, and give it the permissions necessary. And then it will store those permissions locally. Now, this is where I got a very interesting screen. Because if you recall, my only intention with this task was to get it to look at the calendar. And when I gave it permissions, or when I went through the OAuth flow, it asked for this. It asked for the ability to basically see, edit, create and delete everything. Delete, edit, see my files, see my contacts, see my spreadsheets, see my calendar events, see my email. And again, my is it's account. So in theory, this would have been okay. It was kind of like an empty state account. But that being said, I was just trying to do calendar stuff. And so you will see here, I asked, do you really need all these scopes? And it gave me a classic AI URL. Absolutely right. I do not need these scopes. And it reprompted me with that URL for just calendar scope. So if I were to give you a tip, it is watch how and what scope permission you're giving for any of these services. And if you're asking for something specific, only give it scopes for something specific. And if it only needs read access, only give it read access. Just be really thoughtful here. So I just asked for calendar access. No big deal, set it up. And it told me it can do a bunch of stuff. So what did I have it do? Okay, so we just talked back and forth like we were a assistant and its boss. It gave me a summary of what's going on in the upcoming week, what I had today, what I had tomorrow, what was going on this week. And so I gave it a task that I would have normally given an assistant, which is going to the VZero studio this week in San Francisco. I forgot to put it on my calendar. Like I don't remember, can you look it up on the Vercel events page and put it on my calendar. And it couldn't actually find it on the blog and asked me some questions, gave me some options. It did say that I could if I wanted to be, you know, easy and easygoing boss, give it access to Gmail, but I definitely wasn't going to do that. And so after a little bit of back and forth, including some drop Telegram messages, I said, let me give you email access to your own account and I'll forward you emails about it. So again, this is something that I would have done with a EA, I would have just forwarded it and said, can you add this to my calendar? No other context. Now I did have to reauthorize access to its own email. Um, so it went through that OAuth process again, it got the email, it ingested the event details from the email, which was really great, super helpful. It recommended things like adding buffer time for commute before and after, which is definitely what I needed. And I said that I wanted it to add that event to my calendar. Now, if you recall, it doesn't have right access to my work calendar only has right access to its own calendar. And again, it really wanted me to give it edit access to my calendar. And I'm sorry, but absolutely not. And so just like a colleague, just like an EA, instead, I said, Hey, can you just create an event on your calendar and invite me to it and thought I was smart and said it would do that. And it did that really well. So it added separate calendar blocks to my invite. And it was really nice. Now I noticed finally, I found that it was actually on my calendar. And so I ended up at a different time. So I had it delete the duplicate event and actually reset it and it got that completely right. So I would say for a single calendar event, I was a little back and forth. It did pretty well. Like this is a little bit of what an assistant would do. My only complaints on this was actually how it thought about doing it was definitely like, give me access to everything. And I'll just impersonate you and do things on your behalf. And that's really not what I wanted. I wanted it to act like a assistant. So the next thing that I did was I wanted to figure out what more Claude bought could do for me. And so I asked it directly, like, let's figure out how we can work together. I want to stay coordinated on tasks. Tell me how you want to work together. And it gave me some really good options and was pretty flexible about how we could work together. And it called out what it already has, which is calendar access, date memory files, Telegram, where we can communicate, Gmail access, which we just talked about. And here are some options. We could do a to do file. We could use calendar events. We could use email. We could keep notes. What's my preference? And I just said, again, I don't really care how we work with my AI bot. I just said, whatever is easier for you. And then I dumped a bunch of things that are top of mind. Again, this is how I would work with an EA. I just sit down with them, text them, Slack them and say, hey, this was on my mind. Can you get it all organized and work me through it? So what was on my mind? I have an interview with a CEO of Versa. an interview with the CEO of Vercel. I need to reschedule some of our upcoming How I AI episodes, because if you all don't know, I'm coming back from maternity leave and I overbooked myself. I have to stay on top of my enterprise pipeline for chat PRD, so I want it to focus on my CRM. And those are the top priorities I have. And it summarized those priorities back to me, captured them in a to-do, and then started on the first task, which was rescheduling my How I AI recordings and making some recommendations on how I can do my calendar events better. Now, one thing I wanna call out while we're sitting here is this all looks really, really great and super fun. Like, yep, got it, here are your priorities. The reality is one thing that I don't hear people talking about in terms of CloudBot is latency. It is actually real slow. And it's not slow compared to a human necessarily, right? Like if you text a human or Slack an EA and you say, hey, here are my priorities, it's gonna take them a hot minute to kind of organize them, get the work done and get back to you. But when you're used to something like cloud code, like a cursor, like a chat GPT, which is always giving you product kind of progress feedback, it's telling you it's reasoning, it's showing you its tool calls. It's really hard to wait for an asynchronous bot to get back to you on Telegram. I would say that was one of the pieces that has been most frustrating with working with CloudBot is it just feels slow. And I know it's because it's spinning off the sub-agents, it's doing a lot of tasks, it's probably prompted only to get back to you when it has something to do or needs clarification, but it's quite slow. And you'll actually see in the prompting, I asked it, can you always send me an ACK message when I send something, even if you need to research or kick off a sub-agent? Now, it did not do this, so it still remained slow, but I have to figure out how to get it to always respond to me first versus setting off its task. Okay, so back to the task that we were doing at hand. I asked it to give me some recommendations on how I AI podcast reschedule. I had like five in the first week I'm back from mat leave, that is cuckoo belucco. And so what it recommended is that I keep couple episodes, I rescheduled some after Valentine's day. It asked me my thoughts, I gave it some feedback, and it revised its plan. Now, here's where things get fun. Once we aligned on what I wanted to move to later, I asked it to email those two people that I need to reschedule and ask them if they would mind rescheduling to March. I gave it, it's my scheduling link, so they could actually just self reschedule to March. And I said, copy my work email on those emails. And it said drafting those emails now. Now I thought it would draft them. I was wrong, it just sent them. And it sent them in a very funny way. Okay, so then it sent this email, which was lovely. It said, I hope you do well. I wanted to talk to you about our podcast recording. I need to reschedule, except it sent it as me. It sent it as Clairvaux. And it's clearly coming from a separate email address. I gave it a fake name. It was not good at all. And it actually impersonated me. So I actually responded to this lovely podcast guest. And I said, I'm sorry, I'm testing CloudBot. It totally impersonated me and made me sound crazy. But please, can we still reschedule? So thank you to my two guests for being really patient as my AI getting pigs. And I went back to CloudBot. And I said, come on, man, don't impersonate me. You need to reach out as my assistant. I already explained this. I already gave you an identity. But please always identify yourself as an assistant. And it should, I think, knock on wood, store this in its memory and do this in the future. But it was a really funny learning in terms of prompting is really quite important. I thought I was being fairly careful with permissions, which I was. It could only do a couple of things. But I underestimated how much it seems like this tool is biased towards acting as you as opposed to acting as an assistant. And I'll have to look through the repository. And I'll have to kind of get myself familiar with how it's implemented. That's not the intention of this podcast to really understand why that is happening. But prompting really, really matters. And I think the product lesson here that's kind of interesting is, yes, I could have been really, really precious about prompting. I could have said, create a draft of this email to these guests. Send it to me for review before you send it. But at the point that I'm doing that and each turn takes at least a couple of minutes, this is not a productivity tool. This is not making me more efficient than sending that email myself. And so I do think there's this balance between these autonomous agents being user-controlled and being really cautious about how you prompt it and being autonomous and probably doing some things wrong. And I think this is a prompting problem on both sides. It's a prompting problem on the product provider side. It's a prompting problem on the user side. And I don't think enough people are probably sophisticated enough to decompose why one prompt versus the other would do well if you're just a consumer or a prosumer. And so I think this is where a lot of the weird behaviors that you'll see are coming out. So, so far, what have I done with Claude Bodev? I've installed it. I have given an identity. We have rescheduled one event. Or we have scheduled one event. We have given an access to email. We have rescheduled two events now and emailed guests about these events. And then this is where it goes crazy. This is where it gets fun. So I decided to give it edit access to our family calendar. This is a calendar where we have pickups and drop-offs and basketball games and piano practice and my ballet practice and all that stuff. Now, I love this calendar. It's very important to me. And if I needed to nuke it, I definitely could. So I gave it access. And what I wanted it to do was, one, email my husband and I about upcoming week. And get us coordinated on where there were gaps in terms of pickups or conflicts where I was across the city at a Vercel event, and he was needing to pick up the kids for basketball practice. And I wanted it to fill out the rest of my calendar. My kids have started a new basketball season. Our neighbor is picking up the kids on a certain day. All those things, I wanted to get it done. And here is the problem. I gave it a bunch of instructions. And it could read that calendar pretty well. It could categorize the events pretty well. And it had no idea what day it was. And so as I was on Telegram, going back and forth, giving it, can you add this? Can you remove this? Can you change this schedule? I thought it was doing a great job on Telegram because I wasn't really paying super attention. And it was confirming that it did all these things. And then I opened up my calendar. And everything was on the wrong day. I mean, everything was on the wrong day. And if you are a parent, you get this. You're like, wait, wait, wait, wait, wait. Is so-and-so picking up kid number two on Tuesdays or Wednesdays? And I know I moved piano, but I don't think I moved it to that day. So it took me a second to understand the damage it had done. But it had really gotten things wrong. You can see me say, stop. You are setting all these one day late. And it was setting everything one day late. And not only was it setting everything one day late, the CLI tool that it was using to add these events to the calendar could only set one-off calendars. And so it could not set a recurring event. So if I wanted to delete these broken events, I had to go through one by one and delete them. And then the other problem with our crustacean friend here, when you're collaborating with them, is I was on my computer, this one, with my calendar open. It was over here in the CLI with its CLI open. And we were conflicting with each other. So I would try to delete all these bad events. And then it would go put them back, because it thought something got broken. I was trying to add them in. I said, stop. It did not stop because of latency and because of these sub-agents. And so I went through and set up everything correctly. And it went through and deleted all my work. It was terrible. It was really, really stressful. And I said, I had to completely redo. It's like emailing my husband every five seconds. And so it was not great. And it actually never got it right. And I will show and share with you the discussion we had about time zones. But this is another thing that non-software engineers using something like this really have to be aware of is, as I said, on X, the only remaining software engineering problem is time zone conversion. And LLMs just have no sense of space and time. It just does not know when now is. It doesn't have a sense of time passing. Now, I will say CloudBot, because it has these daily files and daily logs, has a little bit more of a temporal sense, but not a great one. And so if you don't understand why a computer could get dates wrong using a tool like this, you're going to get really frustrated. I could at least understand why time zone conversion, maybe there was a UTC timestamp in the Google API. I could at least understand why this was happening and help guide it towards a solution. But it certainly was frustrating and something that I don't think your everyday user would be able to do. So I'm going to entertain you all. And I'm going to tell you, as I was doing this, I took a pause and I took my two youngest kids to Target because we were out of stuff. So I asked if it could discuss things with me via voice. And it said, sure, you can send me voice notes. I can send text back. I could send you voice notes back. Or we could go through Twilio and I could set up a phone call. I just said, let's set up voice notes to your text reply. And so I could press voice on Telegram and have it reply to me as I was on the go. And so while we were in this back and forth on time zones, I want to share with you my delightful voice messages to ClaudeBot because this was a real, real energy. Let's see if we can hear them. OK, so this is me at Target pushing a cart, getting really mad at ClaudeBot. You put it back, but that is a Friday. Friday is current date, so do not change anything. But can you please explain to me why you are getting days mixed up? This league game is on the correct day. Again, please do not change it. But I do not understand why you have the days mixed up. OK, so I am getting super annoyed by this experience of getting days wrong. And it replies, oh my gosh, you are absolutely right. I see the problem now. I was off by one day. Here's all the new dates. And they were still definitely off by one day. So once I sent my mean mom message, it came back with me and said, you are absolutely right. I apologize. Here are the dates right. The issue is, I've been trying. This is very funny. I've been trying to, quote unquote, mentally calculate which day of the week each date falls on. Even though the API is telling me what the date of week is, I should probably trust it. But I was using my LLM brain to decide. And what did I say back to it? Well, I said this. You are a computer. You are not doing anything, quote unquote, mentally. You are making calculations. Can you look in your logs at all and understand where the calculations come from or no? And if you did not enjoy this, that is my very, very new baby crying in the background as I'm lifting him from the car seat into the stroller. It was quite an energy. And again, this is one of those things that, as a software engineer, I get it. I have done time zone conversions for my whole life. I understand that APIs return things in all sorts of formats. I understand LLMs can't do basic math when it comes to dates. It's just too hard. We do not have the technology. And yet, the fact that this model told me it was doing it in his head was so hilarious. So once we had the back and forth about this, it gave itself a rule to follow in terms of getting these dates right. And then I asked it to add it to its rules. Now, the final thing that we did is I asked if it could send me voice notes back. And this is where some of the magic of CloudBot really does come out. One of the things that people have been saying about CloudBot that's so cool is it can give itself skills. It can learn things. It can just do things very magically. And if you're trying to get back and forth voice notes in Telegram, it would have been pretty hard to figure out what API you want to use and what skill and hook it up and use cloud code, all this stuff. And it just did it. So when I said, can you please send me voice notes back, it just sent me a voice note back. So let's see. Yes, I can send voice messages back to you. Let me know if you'd like me to use voice for replies. I can do that anytime you want. That was a pretty magical moment. And I've been giving CloudBot a really hard time in this episode, not because I don't think it's an awesome product. The reality is going back and forth via text with something that has helpful access to your calendar, has helpful access to your email, can learn skills like voice that you can just chit chat to, I actually really liked the form factor of the experience. And I liked the concept of what it could deliver. It was just that the implementation of it had a couple of things. One, too technical for the everyday user. Two, too scary to the security aware user. And three, latency that took some of the magic away from the experience. And so again, I don't think this is a bad product from a capital P product perspective. I'm just not in love with the implementation. And we'll just summarize what I did with CloudBot with my last use case, which is I had CloudBot use its history to create a Next.js app that showed the history of our conversation. And I asked it to redact names, numbers, URLs, email addresses, all that stuff, so I could share it with all of you. So again, kind of a classic AI engineering, AI coding, vibe coding use case. Now, the one thing that I will say is a lot of people are really excited or say they're excited, I don't know if they've used it, to use CloudBot to spin off CloudCode to do coding for them. And where this wasn't the magic use case for me and why I didn't start it is I've been spinning off remote agents with computer access to do coding for me for a while. I use Devon, which has a virtual machine in a local environment and can spin up stuff, access to the web all the time. I use it from Slack, so I can at-mention Devon. I have a Slack bot for chat PRD, so I'm at-mentioning my product manager all the time. Cursor has background agents. Everything has codecs you can kick off online. So I don't know if people are just not using those tools. I guess CloudCode doesn't have one like that quite yet. I don't know if people aren't using those tools, but I've been coding by kicking off an asynchronous teammate, quote unquote, for, you know, two years now. And so that piece was never what I wanted to use CloudBot for, but I thought you gotta vibe code something when you're trying a new agent, and I did that. So what I did is I sent Polly the CloudBot a voice note, and this is the requirements I gave it. Okay, let's use voice from here on out. I want you to document our conversation in a Next.js web app that shows the back and forth of our full conversation from the very beginning today till the end in a UI. I want you to redact anything that is a secret key, a person's name, or a specific place, and I want to toggle between two UI versions of this display. I want you to be able to show me a terminal-style conversation back and forth, similar to a Cloud Code or you CloudBot, C-L-A-W-D-B-O-T, or I want you to show me a telegram-style text back and forth. The content should be in JSON, the same. Again, redact names, emails, dates, et cetera. Replace them with placeholders or redacted blocks, and then generate the Next.js app. I'm gonna use this so I can share this conversation with others without sharing my information or having to do a screen recording. We are eventually going to deploy this to Vercel. Can you let me know when it's deployed to Vercel so I can look at it? So I sent it this message and it kicked off local development, building a Next.js app. Now, when I got back to my laptop that Cloud was running on, one of the things that I noticed is deploying it actually wasn't that simple. CloudBot didn't have a GitHub account. CloudBot I didn't really wanna add to my Vercel account. I didn't wanna log into those things. It seemed like a big rigmarole. And so getting it to deploy without having to set up a bunch of accounts seemed not fun. So what I did instead, don't tell anybody, is I airdropped the repo to my own laptop here. I actually logged into Cloud Code and made some edits. And to be honest, in terms of coding quality and just the back and forth, with the latency of CloudBot and the inability to sort of see what decisions it's making from a coding perspective, I didn't love CloudBot Telegram vibe coding. It's just too slow. The cycles aren't good enough. They aren't incremental enough. It's clearly not like perfectly tuned for the coding use case. It's not like sending me a PR link, all those sorts of things. And so I just preferred working with it on my desktop in Cloud Code and deploying it through my normal system. So that's a little bit feedback there. One thing that I did think was really cool is when I was on the go and it said the app was ready, you know, I was in Target or whatever, and I wasn't at a place where I could run a local machine, it was pretty cool to say, hey, like shoot me a screenshot of what it looks like. And it did. It shot me screenshots of what the app looked like directly in Telegram. So I do think there's some underappreciated aspects, really simple things. Email me that file, share me a screenshot. That are really useful to interface with a laptop or a desktop or a device at home. So I do think this is an underappreciated aspect of being able to chat with your computer. It can do things like send you files, take screenshots, open up browsers. That is pretty cool, especially since we don't store everything in the cloud. All my desktop screenshots are not in the cloud. Some of the PDFs that I download are not in the cloud. And so this was a really kind of like fun use case for chatting with a remote developer. Now, that being said, Devon sends me screenshots all the time. I don't think it's perfect for coding, but it's something to think about. So I wanna end this workflow section with one workflow that I thought it did a particularly good job at. And it was good for two reasons. One, the product interface was what I wanted. I got the full CloudBot bot experience. The second thing was the output was really good. So what did I ask it to do? Well, I asked CloudBot to go on Reddit and research what people would want from chat purity. So I said, go on Reddit. I did this during voice note. I said, go on Reddit, find what people want from chat purity, find what they want from a product AI platform and email me a report. And what did I love about the product experience? One of the killer features of CloudBot is the ability to message anywhere, anything, anyhow. I sent it to voice note, I could shoot it an email, I could text if that was faster and it would reply in kind as text, as voice, whatever. And it would also email me. So it felt very much like an employee that I was working with. Hey, like send them a Slack and they're like, yep, it's in your inbox. Always on, anywhere, anyhow, communication flow for the agent. how communication flow for the agent was really, really nice. The second thing I like from a product perspective is, I've talked about this from a negative point of view, which is the latency is not great. It's just not super responsive and super fast, and it's kind of broken sometimes. But if this is a research task that I don't really think should come back quickly, I don't mind waiting for CloudBot to do a good job. And it did. And it's very similar to an experience with an employee. If I give them sort of a research task or roadmap task, I don't expect it to be returned in 30 seconds, except if they go out, do a bunch of research, and come back to me. And so I wasn't as bothered by the latency here. And then the third thing is, I thought the output was actually quite good. So I'll show you what it sent me, which is it sent me this chat purity Reddit research markdown document, emailed it to my inbox, and it listed out key insights from researching Reddit. And what I thought was awesome is this is right, but it's presented in a really simple, punchy way that I can go action. This is exactly how I would want a PM or a research assistant on my team to come back with insights. And these are the things that we hear. So it was really accurate, integration limitations, both on our side and customer side's hard. No one reads long PRDs. Let's make our PRDs shorter. PRDs need to be living documents. All these things, a couple bullet points, a couple reference links to Reddit threads. And I have a full document that I can go build a roadmap off of. And in fact, that's exactly what I asked it to do. I said, go build a roadmap based on this. Look at our current functionality and tell me what I should build next. This felt pretty magic. I'm probably gonna steal some of these ideas. I'm gonna circle around to that in a little bit in terms of what I think is next for CloudBot. But I do think there is going to be demand, both from a consumer perspective and from an enterprise business perspective on a agent employee that feels like an agent employee. It has a computer. It has account access. It can do things. It does those things well. But I think there are some things we're gonna have to figure out first before we let it loose. And again, you can see this here, I asked it to do a roadmap. It totally just didn't do it. It forgot it, said, let me check on the background agent, and never replied. So again, we're hitting some sharp edges on the product experience. It's not perfect, but it is pretty interesting. So, what have I showed you so far? One, I have told you a little bit at a high level what CloudBot is, although I haven't gone into all the detail about how it works, not really the point of this episode. I've showed how to onboard with CloudBot, including how to connect Telegram to chat back and forth with it on text or voice. I've showed you how I give access to its own Gmail workspace account, as well as its own one password, so it can interact with limited scope. To my data, I've given you some warnings about what you should think about in terms of scope and access there. I've gone through a couple workflows, couple admin workflows, which are simple calendar all the way to advanced calendar management. Did not do well because it doesn't have a good sense of time and space, but hopefully we'll figure that out. As software engineers overall, I asked it to contact partners and guests by email. It did not a great job there because it lost a sense of its own identity. I had it do vibe coding, which it did a fine job at, but is definitely not my favorite tool for AI engineering remotely and asynchronously. And then finally, I showed you my favorite use case was for it to do some complex research and analysis with tools, with web, and it did a really good job and came back to me with something that I really like. Now, one thing that I didn't go check that I'm gonna go check next is, did it teach itself all these skills? Was it really telling me the truth when it said it had rules? A peek under the hood, which I will probably do as a follow-up either on X or here on the podcast, like how does this thing work behind the scenes? That was not the point of this episode. The point of this episode was to show how somebody would come with a blank idea, maybe a fresh Mac mini, install this thing from the command line and actually get it to do things. And I think I showed you it's good at some things, bad at others, and scary across the board. So let's get to my final thoughts here, which are basically that. And I shared this on X, but the whole time I was doing CloudBot, the whole time I was using this, I thought two things. One, this is so scary. This is a terrible idea. Nobody should be doing this. It should not have access to all this stuff on my computer. I should not be sharing these keys locally. I should not let LLMs have access to Gmail OAuth, even if it's a sandbox app. I was like, no, no, no, no, no, no, no. SOS, don't love it. As I said, this is the final boss of security training. You should be very careful about what you give it access to. And one of the things that I'm most concerned about is probably you get the most power from CloudBot if you give it access to your actual inbox, to your actual calendar, to your actual documents, to your actual repositories, your actual GitHub. And I can imagine so many things going wrong with that. Just knowing how it's built, which it's built in an awesome way. I think Pete's done an incredible job. I don't think there's any ill will or malintent in how it's built. It's powerful, it self-learns, it installs skills, it asks for permission. It's pretty independent. All that stuff is great until you have full read-write access to your most personal information. And one of the things that I was thinking as I was preparing for the show is, great, I just gave an autonomous AI agent access to where my kids' basketball practices are. Is that something, do we wanna self-dox to a AI crustacean? Probably not. So I think that's gonna be one of the challenges of this product because the second feeling I had was, boy oh boy, I want this thing. I want AI that I can text. I want AI that does not make it complicated to talk back and forth with voice. I want AI that when I say, hey, can you look at my CRM, doesn't say go to this webpage and press this button and enter this API key and do this and that. I just want it to happen automatically. I want all that. I just don't think this is it yet. This does not feel yet like the interface to get me there. And so I have this real tension between, I think the product from a product experience isn't quite there yet. It's not really for the non-technical, so it really is for tinkerers and hackers. There's a lot of security stuff here that's super scary and can I have it please? And so maybe this is an example of something that from a market category perspective definitely has product market fit. There are gajillions of dollars to make here. I just don't know if this open source YOLO mode terminal tool is the thing. And in fact, I'm gonna take this laptop soon and office space it. For those of you that are very young, that means I'm gonna go hit it with a sledgehammer. But I'm gonna uninstall it. I'm gonna remove those keys. I'm gonna delete the Telegram bot. I don't like this. This makes me nervous. And I'm also gonna go build one for myself. And so I think there is, why there has been such a zeitgeist around this product is it is actually really cool to be able to chat, voice, whatever a very smart self-sufficient agent and hackers see it. And they're also more risk tolerance than the everyday person. But that being said, like husband, please don't connect your Gmail to this. Like mom, absolutely not. Like kids, stay away. Not safe for kids. Like this is just, this is something that unless you have been through a security tabletop exercise and know what to know, I would just be really cautious about how permissive you are in terms of access. And that leads me to my final question of this episode, which is, you know, I think Claude will live in our hearts forever. And in fact, it's probably got a great future in front of it. I love how fast the team is going, Claude bot, but who's gonna build this thing for real? Like who is actually gonna build this thing for real? Who is going to build the consumer version of it? Who is going to build the enterprise version of it? Who is going to get it right? And I think this is a complicated question. And I'm just gonna pose some thoughts as we close out this episode. You know, this should be Google or Microsoft's game to lose. Maybe even Meta's on the consumer side, but this should be Google or Microsoft's game to lose. Like they have the data. They have your Gmail. They have your calendar. They have documents. They have the models. The models are exceptional. They just gotta build the, and you know, they have devices, Android. They just have to build the product experience and have the sort of institutional fortitude and close your eyes legal team to allow some of this to happen. Because I think it's a really cool product built on top of the Google ecosystem. I mean, I think the same on the enterprise from a Microsoft perspective. If Copilot did this, this is pretty incredible. That being said, I don't know if those companies are gonna have the velocity or the bravery to go as YOLO as CloudBot did. So I don't know if we're gonna get there with the big companies. On the flip side, you see CloudBot, open source, great for hackers, but super scary, giving like API key and OAuth access. Smaller companies, startups are gonna see this and want to build this. And I think one of the things that I would warn startups, it's really hard to build on top of these data sources for real, because Google doesn't want to give you read, write, go, do everything access to their data. Microsoft does not. You have to go through these compliance hoops and approvals and reviews. And so while I love the idea of a do everything, do anything bot, it's gonna be complicated from a product builder perspective. It's gonna be complicated from a large company perspective. Who gets the data? And then again, like, is Apple gonna get in this game? This is just what everybody wants from Siri to do. Siri has all your apps, all your access. But again, it's a combination of product building skills and risk tolerance, I think, and willingness to experiment. Maybe Anthropic and OpenAI come in. Maybe we get our OpenAI OS and Workspace tools, and maybe we get Cloud Inbox, and we get some new versions of this. It'll just be really interesting to see how this shakes out. So in conclusion, what are my thoughts about CloudBot? It is scary. It is fun. It does some things really, really well. It is really interesting from a interface perspective, and it doesn't always work. And I'm not sure it's for, I was gonna say everyone, but I'm not sure it's for anyone right now, except for people who are really willing to roll the dice with their AI bot. That being said, if you're willing to do that, the way this has been built, the way it self-discovers skills, the way it stores its memory, the way it gives itself access is really interesting and should inspire a lot of product builders thinking about AI products on what the interface of the future is. I think we are gonna be seeing and hearing a lot more about agents like this. I am gonna be giving you my honest takes about where they are now and where they're gonna be in the future. And in the meantime, I'm gonna go execute Polly the CloudBot. Thanks, and I'll see you next time on How I AI. Thanks so much for watching. If you enjoyed the show, please like and subscribe here on YouTube, or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time.