TanStack Creator: Favorite 2026 Stack, Agentic Workflow, T3 Code & Why JS Isn't for Backends

Tailwind CSS or vanilla CSS? Tailwind. Package manager of choice? PNPN, baby. Favorite JavaScript framework? Solid. If you started a project right now, where would you deploy it? Cloudflare or Netlefy. [music] One JavaScript library that you love but you don't maintain. TypeScript. [laughter] I thought you would answer Zod on this one because of the library being so agnostic there and you can just use Zod if you want to. Zod is a good one. Zod. Valot. Yeah, those are good ones. I I can't live without those. You're right. Thank you for the correction. JavaScript runtime of choice. Bun node ordino node. Did Verscell ever try to hire you? Yeah, [laughter] just out of curiosity. Every single guess I have answers yes to this question. It it's kind of becoming a meme at this point. I'm making I'm making this question all the time cuz everyone says yes. I am curious about your personal agentic workflow at the minute. What is your favorite model? Are you using more than one model? What are you using it right now? I've been dancing back and forth between Opus 4.7 and and the new 5.5. There's a lot of momentum that gathers in some of these projects. Like if you start chats and you start processes or features in one platform, it's difficult to like migrate that to another one. So I kind of have threads going in a lot of these chats. So most recently, I actually was all in on open code for a really long time. I was like doing crazy open code stuff and mixing models inside of the same chat. That was really fun. But I started getting tired of the 2E experience and I tried T3 code which I actually am really bullish on. I think T3 Code is pretty sweet. It's still young. It still has a lot of rough edges but it's very promising. And then Anthropic people that I know convinced me to try Cloud Code Desktop. So, actually the the actual desktop app on Mac, I tried it out and the code panel on it, it's exactly what you would expect. Chats on the left, panel in the middle, you got widgets and stuff on the right, but it got a lot faster and more reliable for me than it than it ever was. I've actually been using that for about a week or two straight, just the desktop app, and it feels weird because I was in the TUI for so long. But similar to that, I've got Codeex opened up next to it and I'm I'm trying to try all the desktop apps this month. I've done that in the past. You know, you go in and try the desktop app and you're like, "Oh, this is not as uh advanced or customizable as the TUI version, but this month I have stayed, which is crazy." Um, so I'm I am running more of the the guey style from Cloud Code and Codeex. And I've got T3 code open too that I'm tinkering with at the moment to try and see what I can get out of it. Yeah, I also have another one that I've been trying out, but it's very early. It's alpha and I actually don't even know if I'm allowed to talk about it, so I'm not going to. But there, let's just say that there's a lot of them out there. I think the interesting one is people are trying to get me into like orchestration, you know, like, oh, you should use this app because it's it's for doing a lot of agents all at once. And orchestration, I mean, do I believe that we're going to get there, but right now the quality that I get out of like agent orchestration is not there because I have to be involved in the process. And so right right now I'm not just trusting like 10 agents to go out and get something done. And to me that just creates more of like a development burden on myself almost. So right now my my workflow is like I have three or four projects going at one time and inside of each project I usually have one or two threads that are active that I'm kind of just managing back and forth. I have multiple opinions of everything you just said. Like uh the first opinion is that I don't like orchestrators yet. I think like they all the sub aents they miss a lot of context in the code quality produced is just not there yet. So I like to go with one session two max and I like to control them go slowly but surely ensure the quality and blah blah blah blah blah. Regarding going to desktop I'm not there yet. So I still love the 2 you know other cloud code CLI or open code. However, recently I found that a real blocker for me is that I was super into cloud opus 4.7 and the real blocker for me was like if you are on the cloud code CLI and the API is down which is happening more than usual you are just stuck like you cannot work literally you literally forgot how to code anymore like you don't you don't know how to type anymore like I literally lost all my m muscle memory you know uh like I know how to review code and I know exactly what good code is but um I just don't know how to type it anymore you know what I mean it's math basically you know how to do it on the calculator but if you want to do it by hand you just forget it right so I've been shifting right now since two days ago actually to open code again I used to be open code user but then I moved it away now I'm back again and I'm happy with JPT 5.5 actually yesterday I had a very productive afternoon which GPT 5.5 and it was super accurate it was super fast like insanely faster than cloud code used to be the opposite actually but now it's just super fast and yeah just happy again happy open code user I guess even though they have a desktop app but desktop apps don't work really for me I guess [laughter] that that is why I'm interested in T3 code for that I like the ability to switch between the models I I know you know there's this whole thing around anthropic and they're really trying to draw a hard line on the moat that they have around claude and I I get it from one perspective but then the agnostic side of me as you as we've talked about is just like no like that should be an implementation det detail and and you should be winning my business based on your model, not based on your ability to lock me into, you know, a specific harness or something like that. So, it it's weird. I I go back and forth and that's why Open Code was so great is because it doesn't happen often. It's maybe every other month, but a new model will come out and you're like, well, I want to switch to that and try it out. And it's really difficult to pick up everything you've done in a in a guey or something and move over there. So if there is a hope it's for something like T3 that is open source that is a guey and that behaves a lot like actually have you tried the open code desktop app? I have not just by the principle that it's a desktop app so you lost me there. I I use it with my son. He's seven years old and I stuck him in front of the desktop app for Open Code and I I installed uh STO I think it's Kitsy's dictation app for Mac and I just said hey son press this button and talk into this textbox and then you can build whatever game you want and he just started building games and the open code desktop app was awesome for that because it's it is still simple very barebones but it's getting better all the time you should try it out we have something called the Laval a boost which literally gives AI skills to let AI know how to work with Laravel with good conventions. How should he write the code? How should he write modern code and etc. In Tenstack, do you offer anything close to this? Yeah. So the the the node world is weird and it's different but there are some similarities there. So where everybody's talking about skills, you want to distribute skills for for your agent know to know how to use it. So let me go through the different layers of AI with you. I'll I'll be really quick here. So the first layer that everyone should be thinking about is type safety. So if you if you're using an agent harness that is worth its that is worth anything. It's going to have LSP integration. It's going to hook right into the language service provider and it's going to know exactly what's happening every tick of the way. Which is why having actual type safety, not just TypeScript, but actual type safety is not just going to make it easier for you to write, but it makes your agent better. It helps like and and not even just safety, but we actually encode the best practices for using all of our libraries into the types so that the types actually steer you into the pit of success. So type safety is the first layer of defense. We can make sure that your agent is doing things correctly before they even compile. The next step after that, I think there's a lot that goes into training data that's becoming less of a thing, but having good documentation is obviously the next thing. And and TANSAC docs are nowhere near perfect. We're always trying to improve that. But something some some of these new things are really interesting to me. So skills are really interesting right now. I don't know if you're part of the npm ecosystem, you're probably aware of something called skills.sh. Really cool idea. I respect them for coming up with this idea and shipping early, but man, I do I think that there is something very wrong with this implementation. And I'll tell you why. Because if if you go to skills.sh and just type in Tanstack, you're going to see a page of skills repos and none of them are from us. None of them are from Tanstack. And even if we did publish our own skills on skills.sh, SH we would now have to compete with the 5K installs of Decar Dearger/Tanstack agent skills tan sex art best practices whoever this guy is he's not the authority this is a provenence question this is a question of trust distribution and registry so what skillssh has done here is they've created a new registry and that registry is going to suffer from the same problems that every other registry has but it's not founded on the same princ principles of something like npm. So where's the versioning? Where's this the security providence? How can you trace that something was published that when it was updated? How do you even update? Right? You've got to use npx skills or whatever. This isn't uh specific to skill set.sh either. It's any skills registry is going to end up re-implementing the same problems that every other registry has ever implemented themselves. So we at TANSAC, we asked ourselves, why would we do that? Why don't we just use npm? We already have npm, so let's just distribute skills through npm. And then the problem was, well, how do you get your agent to find the skills in npm? That doesn't really exist yet. And so that's what that's what we built. If you go to tanstack.com/intent, tanstack intent is a CLI package for both for library authors to make the skills in their npm modules accessible to intent and for normal users to install into their agent harness that will let it look up skills from npm. And you know what's cool is someday agent harnesses might just make this not needed anymore. Maybe they'll just say, "Hey, you know what? we automatically scan your npm modules and look for skills. But today they don't do that. So we built a little CLI to do that for us. And what this does is it means that if you install a Tanstack library and you're using intent, you've got the skills. If we upload new skills through npm, as soon as you upgrade your npm dependencies, you get the new version. And that's the thing is they're versioned and they're they're as secure as you can get for a package manager. you know, you know that the skills you're getting from npm are just as secure as the code you're running from those modules. So, there's a there's an alignment of trust. There are a lot of new ways to get AI to write code really well. And I'm surprised at how well AI does with Tanstack. I know that a majority of it is because of the type safety, but even on top of that, all of the stuff around training data being a thing and and skill distribution is kind of getting solved. I have tons of people now that are oneshotting Tanstack sites. They're oneshotting apps. They're oneshotting migrations from Nex.js. They're building really complex stuff. For the last four or five months, I have not been writing manual code in on on my projects that use Tanstack. Does it mean it's the prettiest? No. But it's type- safe. I know it's going to compile and I know that it's following best practices because they're all encoded into the system. So my confidence levels are pretty high. That's one of the discussions I had like with people regarding like AI in general is like for example if you look at PHP is obviously written on C right it compiles down to C or whatever and uh or assembly or whatever like do we really care about the aesthetics of the compiled code or even the performance like as long as it's like super fast do we really care we just care about the output really you know with Tenstack is probably a little bit different you want to follow like the good principles so it's fast on the browser and blah blah blah uh but I feel like probably for developing this real quick website that you need for providing something to your family or whatever, it's like more than enough to ask AI to just do it with Tanstack. If it's if it compiles good, if the outcome is good, good enough basically. Yeah. And then we want to cover all the bases, right? So like if somebody wants to use Tanstack to build something personal software, yeah, it's good enough. But we also need to make sure now that for the crazy huge companies that are going to be deploying Tanstack at scale over the next 6 months or whatever that it's going to be able to handle the brunt of scale that it's going to be able to handle all of the requests per second during SSR and that when you get into applications that are hundreds or thousands of routes or that you've got, you know, millions of lines of code that the type safety doesn't buckle and that the performance on the server stays really snappy and that page navigations on the client stay really really fast. So I mean there there's a full broad spectrum of performance implications that you need to be aware of and making sure that AI can fall into the pit of performance is really um something I think we should be paying attention to. It's so interesting because we at Laravel we actually solve the same problem as you like the intent thing it's literally Laravel boost that we have it's called Laravel boost does the same stuff versioning skills and blah blah blah blah we also ship conventions to make your code not only doatic but like fast secure blah blah blah blah blah a lot of people in the chat mentioning that it would be really nice to have a lot of 10 stack integration. So at Laval we have this thing we call starter kits and those starter kits work very well with inertia in react, inertia and view and inertia and spelt now as well. It could be potentially a story to explore. I'm going to make sure this that happens. We have a p and then I'm going to show you the p to let you know what you think about uh potentially a lot of uh tensack integration. There are a lot of people in our discord who are who are running inertia as their API as their as their back end. Oh, so it could just work. Yeah. Yeah. They're using Tanac start as the backend for front-end layer. And you know what? That that's an interesting angle because I came from a world where it would be ridiculous to think that you're going to build and scale a a huge back-end system in JavaScript. I'm pretty firmly of the belief that you should just not use JavaScript or TypeScript on the back end. Now, I say that for like real backend workloads. Backend for front end or like this middle ground area. you're writing JavaScript to communicate with your backend APIs and some of that might happen on the server during SSR. They call this kind of a BFF like backend for front end. This is the way that is the way to go forward because I I mean I spent 10 years at a company where all of our stuff on the back end was written in Go. And that's the reality at a lot of these really sophisticated products and businesses is that you're not writing stuff in JavaScript on the back end. And if you are and it works, then that's cool. I respect you. you should probably be using tanstack start for type safety or something like that. But anytime that I hear somebody say, "Hey, I I want to use Laravel or Inertia with Tanstack Start. Can I do that?" And I'm like, "You should do that. You should build your backend in whatever language you think is best for the back end and then when it comes to front end, Tanstack's going to be there for you, but not try to control everything that you do." Have you ever used PHP and Laravel before? Not Laravel. So, I actually started in PHP way back in the day and then and I got really into WordPress and building WordPress plugins and and building custom PHP stuff in school and college and whatnot. But eventually I I got out of that and and you know, the last framework interaction I had in PHP was with Cake PHP uh through some of my co-founders back in the day. I didn't quite make it to to Laravel, but I've kept my tabs and my eyes open on it, and I'm aware enough to know that I I think it's a great piece of software. So, thank you so much for joining me today. If you enjoyed this conversation, hit the like, the subscribe button, drop a comment below on your thoughts about Tenstack. And also, if you want to see an interview I did to Dax Red from Open Code, please go to this video right here. I have discussed a bunch of stuff, including his favorite model, editor, and much more. So, click on this video to watch that right now. That was it for this interview. I love you all and see you guys next time. Boo.