Remote Coding Agents

Welcome to syntax. Today we got a show on remote coding agents and and what are your options there. So what we're talking about is that you of course are probably have used an agent where you're sitting in your editor or sitting in your terminal and you type in the box and you wait for it to come back and and whatnot. And then sometimes you run into the situation where you go, "Oh, I wish I could have these things run when I'm going home, or I wish I could have these things running in parallel on a different machine, or I wish that I did my laptop." I did this the other day where I had an agent running and then I go, "Oh, I got to bring my kid to something." So, I tethered my phone and I kept the laptop open in the passenger seat with my phone tethered to it. And I was like, "This would be way easier if I had a remote setup." Are the same person. I did that same thing on the way to dance. I had the laptop over like like creaked, not like open, but it creaked open. Exactly. Yeah. Just so that it could finish whatever longunning task it was. And I personally have moved so much off of this machine that when I have to wait for something on this machine, I get really annoyed with it. Yes. So we're we're back to thin client days, which is if you don't know what that is when computing and and mainframes were a thing. It was the big idea that like you have a server somewhere running your heavy compute and then your actual like terminal or your your laptop is just like a a window into the stuff happening somewhere else. So there's a lot happening in the space right now. There's a lot of different solutions. There's a lot of DIY solutions and of of course a lot of these big AI players are are trying to um move you to their own solution for this as well. So, we're going to go through all of that. My name is West Boss. This is Scott. What's up, Scott? You ready for this? Yeah, I'm ready for this, man. I'm ready. I'm ready to talk about this stuff. I'm ready for the comments to tell us they don't want to hear about AI. That's what I'm ready for. Scott, the comments really bug Scott. I'm sorry. Uh, everyone be a little nicer to Scott. You can tear into me, though. I I I find them to be either hilarious or um sometimes there's a little nugget of truth in there, and I appreciate that. They do bother me. Why does Why? I don't know why they I've been on YouTube uh since 2012, and I've had more rude comments sent my way than uh many people in this world, and they still bother me for some reason. When I feel it bothers me when I feel misrepresented. That's what bothers me. I hate it when people are like, "Oh, you've gotten the marching orders to talk more about AI or something like, bro, this is just what we're interested in." You know, so much of the the mean stuff is just like out to lunch or they don't understand the full picture because they've gotten that happens a lot with like a short form content clips. Like why not just use X Y and Z? And it's like believe me, I understand the stuff really well and yeah, there's more to it and I can't get this all across in 30 seconds. get upset about people criticizing me if I'm wrong. I'm I'm happy about that. Uh but yes, let's get into it. Remote agents. So, why would you want to run an agent remotely? As we've already mentioned, uh if it's something's running on your your machine, you're going it's stuck on that machine. It's a process that's running on that machine. When something's running on a remote uh location, it could be long running, could be something that you can access anytime, you can access it from anywhere, can access from your phone, from your laptop, from this or that. That gives you a lot of flexibility and a lot of freedom to just have this thing existing, right? Yeah. and a lot of being tethered. Like a lot of people touting the the benefits of remote agents is that like I can I can do something when I'm sitting on the couch at night or I'm standing in in line at whatever or I'm driving my car and I I don't feel like looking at the road and I would like to start off an agent, you know, like there's probably a little bit of work life balance stuff that needs to happen there. But the idea that you can keep stuff going when you're not necessarily just sitting at your desk when an idea strikes you, when a bug comes up when like some of my best ideas for like stuff that features that I should add to a codebase happen when I'm simply just chilling, you know, and and it's nice to like you can make a GitHub issue from your phone or whatever, but you can also just like kick off the work on that thing and then next time you're at your desk or whatever, you can um sit down, you can take a look at at what has been done, especially with some of these longer running agents. Some of these agents are starting to not necessarily get slower there, but they they can do a lot more. The the cursor cloud agents can now use a browser. Um so it'll it'll pop open a browser, it'll visit the thing. Um it'll click on things, it'll check things. and and that's a really cool feature, but it's also really slow. If you don't necessarily care about how long this thing takes cuz you're not going to be working on this project for the next 10 hours, then then they can take their time and and check all the pages and do all your integration tests for you. Yeah. Yeah. I personally really liked this flow. Uh, I know it can feel a little scary to get into some of this stuff, but in practice, I don't I don't know about you, but like Yeah, I kind of I'm kind of liking this uh utility just in general. Yeah, me too. So, like I I'll give you another example. on my personal website, westboss.com, I have a page for all of my tips and and and that is just a collection of usually what I'll do is I'll I'll put together a little video and I'll post it on on Twitter, on Blue Sky, on Tik Tok, on Instagram, on YouTube shorts, sometimes on Reddit, you know, like I post them all over the web. And what I do on my website is a it's a it's a single place where I can bring links to all of those things in as well as as supply code and examples and links to other things that are related to the video. So if somebody's asking me, hey, what was that thing you were talking about XYZ? I have one spot where I can kind of send people to. And I have like I found that process of adding it to my website very annoying to a point where like I I usually just batch them and every couple months I'll go through and try to add them all to there. But recently I made like a a skill that will I can set an agent on it and then I I gave it tools to search all of the different social media platforms for that specific video. Um, and then I also gave it linked to a git repo where I keep the code examples. Um, and basically it's just doing the like grunt work of collecting all the links, collecting the like caption for each one, um, collecting the code for that one, and then and then also taking some additional like I might write two or three sentences that I want to go along with that. And and I I just kind of put that all into a single skill. And now I can simply just like say, "Hey agent, go and add this tip to my website and it will go and do all of that process for me." And and that's really nice that I can just do that from my phone or I can do it on a remote agent and it can just kind of go off and and do all that for you. Yeah, I I love that. We often talk about or people who say, you know what, if if AI could do the stuff I don't want to do, not the stuff that I want to do, then that then that would be like useful. And this is definitely where I I I find a lot of utility in Well, again, it's like the grunt work, the stuff that is just like monotonous, the the things that I don't want to spend any time doing, not the um interesting, fulfilling parts of my my life and career. I I use this type of stuff for for doing research. Um I mentioned before that I use my open claw setup to generate a podcast for me using text to speech where it'll research a topic and then generate that to a podcast. And that's like that's like something I can just say, "Hey, I'm interested in this. Go off and do it." Right. And and that's just the only reason I I choose to have it be in a podcast format is just that's the way I like to to learn things. Not this podcast. For the record, Scott is talking about it like literally generating an audio thing. It's not like researching episodes for syntax. It's not researching for syntax. It's researching things that I want to learn in life and then it creates it to a private RSS feed. So it's not public. I'm not pushing the slop out to other people, but what it's essentially doing is it's distilling blog posts or distilling uh repos or techniques and then putting it into language that makes my brain easier to comprehend it and then it speaks it to me. Uh that it's just like it's using text to speech but also summarizing kind of in that way. And the only reason it's in a podcast format is because it's just uh nice for my phone for for it to exist that way. Um, man. I have it helping uh us do travel planning and stuff right now. I have a little travel planning interface and stuff for it where I can have it go off and research things. Um, and you know, obviously all this stuff requires validation and verification um with AI. But yeah, just in general, go go back and listen to our episode on OpenClaw where we go pretty deep on some of the things that I I use these things for. And and this episode won't be about OpenClaw specifically, but most of the time when I'm using remote agents, I am using them via OpenClaw just because that's what I have set up. Awesome. So, we're going to talk mostly about like like coding remote agents, but like like Scott says in even with my stuff is like this stuff will very slowly or or very quickly move into stuff that is past coding as well because you realize ah yes, this is nice for writing code and and submitting pull requests and whatnot, fixing bugs, but it's also there are some areas like like Scott said, travel planning that it can be useful for. All right. So, the way we're going to break this down is they're going to talk about when they will actually run. Um, uh, we're going to talk about different environments and access to things that they need, where they they run. So, like, are you hosting a server? Are you using some sort of service? Are you running them on your laptop, an extra older laptop that you have plugged in in your office? Um, are you using like cloud code, open code? Are you using cursor agents? Are you using some self-hosted thing? There's a lot of different options there. So first of all when they run so of course they can run when you prompt it right that's that's the very simple example but also I think interesting space where they run is when when something happens when an email comes in when a web hook is hit um one big one that we're we're running on the Mad CSS website right now is we're using Sentry so when a new error is detected in Sentry uh we have it automatically I I don't I don't even know if you know this or not, but I don't this Yeah. In just I just went into the Sentry dashboard and hooked up my cursor API key and and what happens is that when a new error is logged in Sentry, it then takes that error and all of the metadata around it and then paste it into a cursor prompt for you and it runs a cursor cloud agent. Um, and then cursor cloud agent has your codebase. it has access to all of the error information from our Sentry and then it it it will look at it and see if it can figure out how to actually solve this type of thing, right? And it will also have Sentry will give it a whole bunch a whole prompt as well from uh their Seir product as well. And and then like if it can find it, you can set it up to to submit a pull request and automatically do that. So like it's I I had a joke the other day on Twitter which is like try claude which is essentially like try this code and if not you catch it and just that you just send that info to the AI and that that's essentially the setup that we have right now is if there were new errors that were to happen logged in sentry and they automatically can get fixed which is a wild world. It's a wild world. Self-healing software really yes is what that is. Uh that's really neat. Um Also, like you said, speaking of Sentry, Sentry also has the ability to man like uh monitor agents too. So, you can have AI monitoring via Sentry. So, if you are running a lot of agents and you want to make sure that they're being effective, you want to trace and debug what they're actually doing, their tool calls and all that kind of stuff, man, you can use Sentry for that. So, like they always are innovating this product, which is one thing I really love about Sentry. Head on over to centry.io/ io/sintax. Sign up using the coupon code tasty treat all lowercase all one word if you want to get two months for free. But man, there's so much cool stuff in here and I really need to get in on this agent monitoring stuff. I I think this looks really super neat. The LLM call account or the the tokens used is really interesting as well because some of the stuff is like, was that worth it? You know, um I spent $11 to fix a typo on the website. Was that change background color to red? Yeah, exactly. Um well, now now you can you can see the token usage, right? And put a dollar sign to to that type of stuff and say, "Yeah, it was use it." Um so that's kind of like when they run. You know, it can be triggered. You can be done via a prompt. Let's talk about where they actually run because there's so many different spots that you can run it. At the end of the day, you need either a computer or a server running somewhere that can can sit there and let this agent sort of churn away, send API calls back and forth to the LLMs, take a look at your code, you know, all of that type of stuff. So, this could be a machine that lives at your house. I know a lot of people have taken a Mac Mini. I personally have a I'm pointing to the ground right now. I have an old MacBook that I have set up to never go to sleep. and it's it's just hardwired into it. I'll probably throw it into my server rack at some point and and that is allows me just to have something running constantly. You have a Mac Mini, right? Yeah. Yeah. It's so funny that became so trendy. I bought this Mac Mini like what two years ago or something because my Naz exists, but the Naz hard hardware is like garbage for a lot of stuff. Like I wanted to do video transcoding and stuff that the NAS couldn't handle. So I was like, "Oh, I'm going to get a Mac Mini. I'm going to run my software on the Mac Mini and containers and then just use the NAS as actual drives instead." And I've been doing that for a little while. And then now I run a lot of dev processes off that machine. It's not powerful enough because running a Minecraft server for a bunch of kids. I'm running Open Claw in there and I am running uh a bunch of dev process. So, when I can upgrade this, I will be getting a Mac Studio to throw in here. And I'm also going to get a new laptop at some point in here, cuz this thing is starting to chug. And I got to just say, I'll probably just end up tossing uh this this bad boy on there as well, just as a long running machine, considering this thing has like 64 gigs of RAM, and that that has like 8 gigs of RAM, the Mac Mini does. So, yeah, the Mac Mini is is not powerful at all, but it doesn't need to be for what it's doing. Yeah, just running API calls back and forth. I I found the same thing as well. I have a Synology and I'm running I don't know probably six or seven different Docker containers on there. Um, and there was a couple that I needed to to kill because I started to run out of resources on the Synology. Um, and that's why when I'm starting to do this like remote agent stuff, I was like, "All right, I'm going to throw this stuff on like a on a MacBook Pro, older MacBook Pro." But you don't need like an an older MacBook Pro. It's simply a Linux box or something like that. It can you can just go to like one of these computer wholesalers, Raspberry Pi. Uh maybe like I I know people are doing it, but you you I guess it depends on what you are doing, but you probably want a little bit of more of a ump behind it. A lot of people are also just like simply just going and buying like a 510 $15 a month um VPS. it's just a Linux box instead of having it live at your home. Like you might not have very good stable internet or power. Um you you want that running on like a server farm somewhere. So you can do that as well. Yeah. I will say uh something that people don't consider often uh refurbished. I bought my Mac Mini as a refurbished Mac Mini. Uh you can buy I don't know. I know Mac Mini there might be constraint right now constrained right now in terms of like what's available but when I bought mine it was refurbished so it was considerably cheaper. So uh just something to to keep in mind. You don't got to you don't got to shell out big cash for it. That's for sure. True. You can go buy some like Dell like some Dell like box that you know you go to these computer stores in like a plaza and they have like a stack of them like 80 high 200 bucks something like that. That'll that'll work as well. MicroEnter. If you got a microenter near you, you just pop on over to MicroEnter. CJ just bought a little little machine with like an integrated uh memory and stuff on there. It rips. Oh, that's great. So, that's like like a like a self-hosting route. Um, if you don't want to go the self-hosting route, there are lots of obviously good paid options as well. Cursor has this thing called cloud agents which you you can kick them off from your like cursor desktop and you can say run in the cloud. Um and then they also have this new thing called like long running agents which you must kick off from cursor.com. Um, and they will run, I don't know, you can set like five, 10 until done. Um, like hours that to have these things run and they will just run as as long as they they need to. So that's kind of cool because you can just like kick off stuff. You give it access to your your git repo. Um, and then when you're in a cloud agent, they they essentially give you a little abundu box. um they give you a terminal and then they also it's it's running like a VNC so it also has a browser built into it which is really nifty and if you go into like the if you go to cursor.com and log in you can like literally use the browser if you want um I installed claude code on my cursor box just to see if it would work. Um it does it it worked just fine which is like it's not some like like paired down environment. There's tons of memory in it. I I checked it. It's like 18 gigs of memory or something like these are pretty beefy boxes that that you're getting there. And then what that will do is it will just run it of course in the cloud. Um but then it also has access to like a terminal if it needs to run some terminal commands and has access to a full-blown uh Chrome web browser. So it can it can surf it, take screenshots, get the DOM tree, see if elements are on the page, um all of that kind of cool stuff which is is really neat. I think the browser is an essential part of this. So yes, um the one thing about the cloud agents is that is the like the the story of bringing it local. Um there if you're using the longunning agents I I ran into this this weird situation where I would kick one off on cursor.com and then it would submit a pull request and I go ah it's not really there. I want to I want to work on the code now. So then you have to like get pull the like transition from cursor.com to like the cursor editor is not as smooth as I would like it to be, but I imagine that will probably get better. Yeah, which is one of the reasons why I kind of like the local flow that I've established because that handoff is is really seamless. I think that like that that is a a major a major pain point for me personally. Like I don't want to prompt into Infinity when I can make a a change myself and and just wrap something up. Yeah. Yeah. Exactly. That's that's honestly why I've liked a lot of these like editor like I'm like I've I've been using Open Code, Cloud Code, and Cursor. Um but for stuff that like I actually care about the code, I find myself going a little bit more towards cursor simply just because you you have that I know that you can open. Everyone's always like, you're going to open it in an editor, you know, and I cloud code will has a has a VS Code extension and all that type of stuff. But I just run I run the two u two gooies. I I run the open code guey next to my Z guey and I tab back and forth like a caveman like a madman. Okay, so let's talk about that. Then the next kind of option you have here is just like a CLI on a box, right? you you SSH into your box, whether that's a Mac Mini in your office or whether that's like a a buntu that you've rented from some BP VPS provider. Um, and then a lot of people are simply just sshing into these things and then running the CLI of their choice, right? Cloud Code, Copilot, Cursor CLI, Open Code, there's probably 80 other CLI products out there. Um, and you simply just use the terminal to just type in what you want. Um, and then you get push or whatever or you have the agent itself do that. And then a lot of people additionally have either created UIs where like it's like a simple little server that they can instead of having to use the terminal SSH from their phone, it just makes a little website where where you can go into it. Open code has open code web which you can just type it in and it spins up a little server and you get a nice guey for it. Cloud code is rolling out this thing called remote control where you can you can just run slash remote control and then it will spin up a little server and then you can use that um from from your phone on like the cloud claude app or something like that. I'm curious what your solution is there Scott. Oh yeah. I I use um open code open code web uh I'm running it on any place that I have dev processes whether that's my laptop or this the Mac mini. Um reason being is if I'm running open code web rather than the open code guey or the open code 2 on here I just pop to that URL that it serves on my phone. Um and then I can prompt it no matter what. And so that way I have multiple different open code sessions on my phone for for whatever laptop or or machine it's working on, but I have it all connected via tail scale and everything has a tail scale domain or an IP associated with it. So that way that way I can just access it on my phone even if I'm off network because I'm still on my tail scale network. So the tail scale piece of it for me is massive. If you want to see how I do that, I have a video on the YouTube channel about how I run that setup uh that details exactly what's going on there. But because all of these uh systems are all running on the same tail scale network, they're all running open code web. I can access any of the sessions from any of the machines. And then I can also SSH uh to remote access into the code anyways if I want to edit the code, which I do via Zed. Yeah. So, so if you're working in your office and you want to like let's say you want to fix something on the the syntax website your process would like what's your start to finish process of you're sitting at your laptop and you want to do a feature on the syntax website with this these tools. So the syntax site specifically is still running on this machine. It's like the newer stuff I'm I'm working on I do put on there. I because I have so many things that I've like the database stuff. I don't want to transfer it all over to the Mac Mini. So, if it's something that I had been working prior on my laptop, it's still on my laptop for now. But I use I open it up in Zed. I got my zed here and then other monitor I have open code web running. And open code web the it's a guey. And the way that works is that it doesn't need to run for each individual project. You're not running like eight instances of open code web. if you have all these different setups, right? Not like the twoe version where you run open code on a directory and then you have to have a bunch of them running at once. The the Open Code web is basically just a a guey where you can add as many projects as you want regardless of where they're living. And so I just have them loaded up in that sidebar there and I can just click on it if I want to interact with it via Open Code. Mhm. That's it though. Yeah, that that's great. And we'll talk about environments in just a second as well because like that's that's half the battle is like getting your your environment set up with all your databases and API keys, all of that stuff, right? It's my least favorite thing. Yeah. Can be a bit of a pain. Yeah. Next one we have here is the entire editor on a box. Um and honestly, I think that this might be where I want to move with a lot of this stuff as well. Um, I don't think that there's a great solution here just yet, but running everything. So, like the last project I did, I used I I have not my laptop. I had that the second laptop and then I use a SSH plugin to remote into it. Right? So, on VS Code, it's called VS Code Remote SSH. Cursor has one called cursor remote SSH. And essentially, you're still using your editor on your computer, but the the processes and all of the code, much like Scott setup, is running on that remote thing and and any terminals you spin up and whatnot are all running on that remote machine as well. Um, GitHub code spaces kind of takes us one one step further and same with like this. There's a there's a something called code server which allows you to run literally run VS code on the machine and then you can just visit it in the browser which honestly I think like that's like where I'm going with this type of stuff because like I want to just type in a URL and have my entire development environment. Um, I think I think there there are there's a lot of benefits to that. Uh, especially as our laptops are aging here. It takes some of that processing power off the laptop. Um, that helps your battery last longer. Um, there are some negatives if you're on like a bad Wi-Fi area and stuff like that. Um, I don't find latency to be a huge problem in my current setup. But again, I I I guess I, you know, working on an airplane might be a bit of a different situation or, you know, we'll see when I start to take this setup a bit more on the road as I I go here. Um, but I really do like this setup and and it has the added benefit of being able to test everything on your phone, especially if your your everything is running on another machine. You can access it on your tail scale network if that's what you're choosing to use. I can then use my dev. Not that you couldn't do this without this, but I I'm able to really easily with a domain and a port just easily just use it on my phone and try it out and it's like the dev site and I can type and I can see I can get the live reloading right in my phone and everything like that. It's it's just a really nice way to work. I would I would say I I did this as well. I use Cloudflare's um Cloudflare tunnels with their what is it called? Their access product, which is similar to Tailscale, but probably way harder to set up. And I would say if you're going to do any of this, get that stuff set up right away because as soon as you start like connecting to IP addresses and then IP address changes, um that stuff is such a pain in the butt to to get working first. So like just get it out of the way. get a nice domain for whatever it is you're working on and then that way it doesn't matter if you're on your local network or not and you can just go ahead and use it. Also, uh standardized ports um like per project. The way I do it is I have each project that I have I set the ports to be that project's name in lead. Then I'm always going to remember what the port is of any given project. I'm not like, "Oh, is it 5173 or 5174 or 5175 or or any of that stuff." I I've been standardizing on the lead for the name of title of the project. That drives me nuts. Especially like these AI agents will spin up like a local host 3000 and then it's like, "Oh, something's running. I'll do 3001." It's like, "Yeah, you started it on 30,000." And then like I got to the point yesterday where it's on port 3007 and I was like, "Okay, you you've got seven feet processes running and you can't keep track of them all." Um, oh my god. I'm I I'm at a point where I think I need to just like do a strict port. Um, I I always do unique ports for most of my existing projects cuz I hate it when my like browser history is like suggesting things that don't exist for that project because you've used that port. or if you've got like um like local uh cookies or if you've got like a service worker that's been registered to that that's all a pain in the butt. Yeah, I know it's a it's a massive it gets tough folks. Uh yeah. So yeah, similar to cursor's cloud agents, cloud code also has this concept of cloud agents where you can just like log into the website, you can log into the cloud code app, you can use the desktop app. Um and if it has access to your your git repos, you can you can kick stuff off and and sort of let it rip. It doesn't have browser access which is one of the things that cursor has over it at the moment but other than that as far as I can tell there are relatively similar. Yeah. Open code. Did you see this that open code is launching some sort of like orchestrator um for if you're running like local code on your computer or sorry open code on your computer and you're also running on a remote server and then you also have it in like a work tree and another one is a container. They're going to have like a single interface for managing all of these things, which seems really nice. Huh. No, I didn't see. It's so funny. You said, "Have you seen this?" And I'm thinking, "How could I have missed this?" And it's like, "It came out yesterday at 4:00 p.m. So, it's like, oh, yeah, no wonder I missed this." Interesting. Just came out and it's all it's built by James Long. He used to work for Stripe for the longest time and he's a huge local first guy and he says a lot of the like CRDT stuff that he he used in previous projects is going to be work on here. So kind of interesting. I'm I'm excited to to try it out. It's not out yet. So he mentions actual budget. Actual budget was one of the very first projects that got me interested in local first stuff. FL um really interesting project and uh yeah so he he's been involved in this stuff in the right capacities for a long time here and uh he's definitely somebody that I I pay attention to for sure DIY agents so I we asked on Twitter like what are people using right and a lot of people are saying like cloud code on a box I SSH into it open code web um but a lot of people said I I just use a DIY agent so um Pi we had the creator of Pi on here as well this is a um basically an an agent that you can just let let her rip and and sort of make your own cloud code based on that. And that is what OpenClaw is is built upon as well. And a lot of people say I simply just use OpenClaw and it has access to my my cloud code and it can just run stuff for me. Yeah, mine does not have access to my clawed code or anything like that, but I I do I use my open claw way less for coding than I do for personal assistant style tasks because I just like my flow of working in open code. I agree. The chat interface of OpenClaw for doing coding projects is awful. Oh, I don't even use their chat interface at all. It was It's No, I I went and I I ended up building my own chat interface for it as well, but it's just not that's not the interface for for building this type of stuff. It's not not great. I much prefer to use like like a purpose-built um agent orchestrator for that. Yeah, same. Other interfaces mentioning these these are not just like like we we of course talked about like you have an editor interface, you have a web interface, you have like an iPhone app interface, but there's there's other interfaces for this thing as well, right? You can mention in a pull request is very popular for a lot of these things. You can mention it in Slack. Um, you can have things that are sort of like listening to your conversations. You put it in a in a chat room. I know modem.dev is is working on that as well where it will simply just like listen to what's going on in your GitHub issues or your Slack chat or whatever. And then instead of you telling it to do something, it will just, oh, I noticed people are talking about this. Maybe I'll I'll I'll start working on that. Um, I can get down with with the listening thing like having it listen to conversations in GitHub Slack chat. Like I could get down with that, but I you will never catch me using Slack as my interface for coding in I don't even want to be in Slack uh in general, let alone telling it to do PRs and stuff like that. You miss me with that one. Yeah, no thanks. Um, next one is Warp Oz. I haven't used Warp Oz beyond us just looking at their really nice landing page. I do use Warp fulltime though as my shell. I've moved back to Warp ever since we had that conversation. But Oz is the orchestration platform for cloud agents. Spin up parallel cloud coding agents. It seems neat. There's just so many competing things for my time and energy right now. I haven't picked this up. That's the thing is that there's so much cool stuff out there that like there's only so much. That's that's probably the hardest thing for a lot of these companies is simply just getting people to try it. You know, you got to connect your GitHub, you got to get the app, and like breaking people's flow is is really tricky. Um, but this is kind of the the way that they're going with this type of thing, which is is kind of interesting. I kind of frustrated that a lot of the stuff clutters the terminal that I know and love. Um Yeah. Yeah. But I think that this is kind of a an interesting move as well. And there there's several other bazillion other people working on different orchestration platforms as well. I'm going to say I think I'm out on cloud for this. For me cloud like which is like some existing service like cursor cloud cloud code cloud. Um you're going to you're running it all yourself. I I want to run myself. I want to run my stuff like here. It doesn't I want to run my stuff at my house. That's it. Yeah, I want to run it on my house. That's really just it. I I don't feel like I need to run in things on the cloud. I don't necessarily see the benefit to that. Uh especially with my tail scale setup, I can access it from anywhere. What's cloud getting me that I'm not getting from running this on my own tail scale network? No, I think I think if you're apt enough to run it yourself and to set all that stuff up, I don't I don't think there's a major benefit there. Um, I think what is going to get me is whoever can figure out the transitioning thing the best. Meaning that I start something on my laptop and then I can open it up on my phone or I can add it in in GitHub and like it's all just works well and that when it comes to the time where I need to edit the code, I'm not doing any weird pulling locally and and merging in their changes, you know. I just want I just want a folder that I can open and and continue it on whatever platform that I'm working on. Yeah, I feel that. That's exactly same for me. All right. The other thing we want to talk about here is the environment that which they run in, right? Because the the reason why people love to run it on their laptop is because that's where everything is set up for their their dev environment, right? That's where I got my stuff. Yeah. That's where your stuff is. That's where it's where you want to run it, right? So, what what type of access does does an agent need? You need like a fully specked out terminal. You need like a a Linux box or a Mac or Windows that's running somewhere for it to run. Um, it needs access to your your GitHub. It needs so that you it can it can branch out. It can make workspaces. It can do pull requests. All of that that type of stuff for you, which is is kind of really nice. Yeah. And depending on what you need, it might need to have access to various local tools. Like I I do a lot with ffmpeg or or sharp to resize images. Like if it's going off and working on media, this thing needs to be able to access those tools. Um whatever those may be. Any kind of CLI really. Uh everybody's coming out with a CLI now, too. Like Sentry has a CLI that's great that I use all the time for my Sentry stuff. Um, Obsidian has a CLI. These CLIs are are tools and bins or whatever that the AI needs to have access to. Yeah, I the other day on the Mad CSS website, I had a whole bunch of images which were like way too large. Um, and normally what I would do is I would like write some script where it would just like resize and compress them or like maybe maybe install something like a V plugin that would like resize them on demand. But I simply just like typed in to the box. So I was like, "Hey, can you resize these to be max 500 pixels wide, max 800 pixels high, and then compress them 80%." And it just did it all for you. Um, which is great. And it had access to all of my like local CLI tooling, which you may have to then install and set up on some of these these other environments. Um, it might need a browser. This this is kind of a big one as well, is like if it needs to do things via the browser. Um whether that is access things that don't have an API, you know, that's how a lot of these bots are are are working right now where they're simply just using the browser as if they were a human, whether that's like loading it up, loading up your web page in the browser and and inspect element and whatnot. Um browser is a big part of that. Um and there's only so much that you can like spoof and replicate via CLI calls. So, I think that that's that's a big one and that's a that's a big like thumbs up for running your own, especially a lot of the bot detection right now. If you're if if you're like coming from like a data center, if your IP is showing that you're like a request is coming from some data center, it's very high that your your request is going to be blocked because it's it's either somebody running a VPN or it's some script that's running on a digital ocean box somewhere, right? But if you're if it's coming from like the IP address of like your home where where you're also doing legitimate stuff, you know, your your daily work, there's some points that that are awarded there. I like that. Uh that's one of the things I like about OpenClaw having that just baked in uh is that like if my wife is giving it a URL and like asking it to research or whatever like that, it's just straight up opening it on the browser on the computer as if it was me doing it on the browser on the computer. Again, for all those reasons, that makes a lot of sense. Um API keys. This is probably the big one. you need to give it access to your yourv file, right? Like whether that's database credentials, um whether that's like API keys for different services that you need to connect to um deployment. Like one annoying one that I want to get set up is that I was prompting cursor and then it would say I've finished and then it would send the pull request and then my Cloudflare workers would go oh a pull request and then it built it and then and then it would break um like the the deploy would break and then there was like there was no connection from the the logs of Cloudflare back to the agent. Um and I know you can do that. I just didn't set it up. But ideally, what it would do is say, "Hey, agent, check back in 4 minutes or whenever this thing is is finished building and look at the logs. Did it did it break?" You know, and if so, come back, take the logs, stick them back in in your prompt, try to fix it again, send a send a a pull request again. It will build. And you just do that cycle forever and ever until it until it actually until the monkeys write Shakespeare, right? Until the monkeys write Shakespeare. Yes. Yeah. Yeah. Absolutely. Um uh web search. I haven't really figured out web search here, Wes. I know that um you can pay for a Brave web search API token. Um and I just don't know if I'm ready to do that. Let's see how much it costs here. It's $5 per 1,000 requests for search. Man, that's that's a lot because I'm not sure I'm ready to do that. Yeah. How does Open Code do web search? Cuz when I'm using Cursor or Claude and it's like search the web for 90 million things and and scooped them all up, you know? And I'm I'm wondering that they must either have agreements with Google. Agreements. Yeah. Or and like that can't be cheap, man. That cannot be cheap. But yeah, web search is is a big one. So, what does Open Code do with with web search? You have to set it up, right? I've never set up web search and I almost never don't ask it to do web search. I'm always giving it URLs, but that's really So, you do the you do the web search like a I do the web search. Let's see. Um, XAI web search API. What's the pricing on this one? For a lot of stuff like your MCP servers will cover you, right? You you got your Cloudflare MCP server, you got your contact 7 MCP server. you know, that covers a lot of your docs and whatnot, but a lot of times it's like this is a weird error web search. Oh, here's somebody else that had this this issue in a GitHub issue. Here's how they fixed it. Let me apply that. See, I do sentry sear root cause and then the AI usually can handle it from there. Boom. Yeah. And I bet that will also get harder as well. You know, like if Google wants to like clamp down on this and make their Gemini product even better. I wonder if they're going to be like, "Right, no more web search for anybody else." That's that's assuming that they are doing that. I wonder what's going on here. I haven't found a solution to this. Folks, if you have a best solution beyond just using the Brave API and paying five bucks for every 1,000 search queries, let us know. I'm very curious about that. Open code uses XAI. Yes, XAI. $7 per every 10,000 searches. That's not cheap because it does I don't know like a a single prompt could could do seven, eight, nine, 10 prompts or web searches at once. So, you probably blow through that pretty quickly. Yeah, it does say for free. What is the free? Because Google's AI tool is saying that you can get up to 10,000 web searches for free with Exa. But that's Google's AI and we all know that that thing lies like crazy. Is XA from Google? No, no, no. I'm just saying I I was googling it. Um, run up to 1,000 requests for free each month. That seems like the answer. 1,000 for free each month with Exo. And they got a great looking design on this website. So, I think I could do a free tier on this bad boy and then just if I hit that a thousand say, "All right, well, never mind." I think I think what a lot of people are doing is they're just the they spin up a browser in the background and then they they are consuming the output from that which is like goes to show like a lot of people are saying oh my my website traffic has either tanked because no one's searching for their stuff anymore or has exploded because of the it's these agents visiting their website. I would like to just build in something like that for my like a little search skill that just does that just because just you just crawl the web yourself. Why why are we doing using Google? We could just vibe that in the Put an agent on it. Yes. Yeah. Throw an agent on it. Why do we even need these? We could just throw an agent on it. Yeah. Yeah, that's what we need. Yeah. Have you tried? I have not tried wealth but I understand the concept of you it just you give it a list of things you want it to do and it just loops through those things one by one until it's done essentially yeah it's basically a bash loop and then it like fixes and repairs and tries and lessons learned it's I found it to be garbage but um I'm not an expert by any means on it there are smart people who are saying that it's effective for them but I I did it a few times and I was like I don't need this in in my life here I like more control I think than that is looping over something 20 times until it gets an answer or more. Have you one thing I did the other day is I was working with the cursor API and for some reason in the cursor API I was getting the actual system prompts in the API and stuff that they were not showing you in the like website. And uh it it was doing things like um what was it? I forget what it was, but you you could it leaked a little bit of the prompt there and I thought that was cool. And then I was like, I wonder what other system prompts you can look for. And I found all the prompts in the open code repo as well. And uh their main prompt is simply just like keep looping through this list of to-dos until there is nothing left uh to do. And it it's actually kind of funny reading the prompt and you realize like, oh, a major part of this software, obviously there's a lot to it, but a major part of the software is simply just these seven or eight paragraphs of instructing the AI how to do something. No bugs, please. Yes. No mistakes. No bugs. Yeah. Keep going until you uh have nothing left to do. That is remote coding agents. Let us know what your setup is. I think a lot of people are just figuring it out right now and and trying to get everything dialed in, but I think it will be an interesting space in the next couple months. So, leave a comment down below what your setup is. I would love to hear it. Yeah, let us know what you're using, what's been effective for you, what has not been effective for you. Um, and um, yeah, that's all I got. Wes, you got anything else for today? That's it. Peace. Peace.