T3 Code broke me

0:00All the way back at the start of 2026, Julius and I got really hyped on the new Codeex app by OpenAI. It blew us away. It really felt like a different way to build. There were ideas I had seen other places, but it felt like the right combination of things, which is why we were really upset when the performance started to get bad, and as much as we would bother the team, we just wanted to

0:19access the code and fix it ourselves. The OpenAI guys, as awesome as they were, wouldn't give us access to the source. I was hoping they would open source it, but they chose not to. On top of that, Anthropic's Cloud Code desktop app just wasn't particularly great. And all of this led to us exploring other options, concluding we didn't like any of them, and we decided to build our own open- source alternative. It's called T3 Code. You've probably seen it in a few of my videos. And I'm going to be incredibly honest with you guys. Despite being open- source and built by incredibly talented devs like Julius and

0:48Maria and also me for many many months, T3 code was just the open-source worse alternative to the claude code and Codex desktop apps. It absolutely had its benefits. Like you could throw your codec sub, your claude sub, your Grock sub, your cursor sub, and more in one single place, which was genuinely really nice and convenient, but you had to accept that the quality of experience you got was worse because it's being made by three people part-time as an open source project and not by a trillion plus dollar company with thousands of engineers working on the project. I'm going to be incredibly real and say I was almost a bit ashamed to

1:24push it because I knew what we built wasn't as good as the alternatives. I was using the Codeex desktop app as much if not more than I was using T3 Code. But then things started to change. Julius went hard on this project and pushed it way further than I ever thought it would go. I started coming in and making UI and UX changes here and there. We decided to do a bunch of stuff that the labs weren't interested in, like changing how the sidebar works, changing how you control things remotely, changing how your subscriptions integrate and work with each other, and all of these other types

1:54of things. And around July of this year, things really felt like they were working, and we no longer were just a crappy open-source alternative. Instead, it felt like we had something unique. And you can see that in our numbers, the very quick ramp up and now we're about to hit 550,000 users and 140,000 weekly activives. It's honestly kind of surreal. And it made me realize I haven't talked about T3 code enough. It shows up in my videos here and there, but I haven't really explained what it is, why it's growing so fast, what makes it different, and generally speaking, all the things that make T3 Code as cool as it is. And

2:36rather than continuing to tease these things here and there and rely on the rest of tech YouTube to cover it for me, which by the way, shout out all the other tech YouTubers who have been making videos about T3 Code. It's helped a ton. I decided it was finally time to sit down and show all of it. give you a real overview of not just what T3 Code is, but how it's changed, how I build with it, and why over half a million developers are choosing to work with it now. So, if you tried out T3 Code back when it was this ugly, gray, sloppy mess, definitely stay tuned because

3:06things have changed a lot since, and I can't wait to show off what we have done. But, as I mentioned before, T3 Code is open source, requires you to bring your own sub, and we don't make any money off it. In fact, it's quite expensive for us to run. So, I hope you can forgive me for a quick break for today's sponsor. Agents can do pretty much any task as long as they have the tools they need to do it. Sadly, not all tools are created equal, and there's a lot of things you can't do. Even if you give an agent a really good browser, there's certain data it's not going to

3:32be able to access. However, if you give your agent Appy, it can access pretty much everything you'd ever want and more. Appify provides tools for agents, over 90,000 of them, and these tools allow your agents to get data from places they otherwise probably can't. They have one of the best MCP servers I've ever seen. Normally, these are full of bloat and really weird syntax. But the Appify MCP was trivial to set up and isn't wasting a bunch of tokens like all the others do. I added it to Cloud Code, so now it works in T3 code. And I was able to ask it questions like, "What are

3:58the features that people are asking for? What problems are they running into in my repo?" Normally, I'd only be able to access GitHub for this, and even then, I'd probably hit my rate limits pretty fast. By using Appify, I'm able to scan many other sources. It's honestly a relief to see a company using MCP correctly after all of the messes that I've seen over the last few years. This is a thing I see myself using all the time, especially when I need to get data from sources like YouTube, Reddit, HN, and more. Make sure your agents have the tools they need today at

4:25soyv.link/appify. It's hard to summarize what T3 Code is, and I almost feel like a bad founder for it because it's not just like one simple thing. The easiest way to explain it is that it's an open source alternative to the Codeex app that supports all your subs. I find that this is the best way

4:41to convince people to at least try it. But this is not how I think of T3 code at all, and it's certainly not the reasons that I'm using it. I put together a long list of all the things I love with T3 Code alongside my chat. Thank you all for hanging out here. I'm going to do my best to try and summarize like the top level three and then we'll dive into the details. So, the first key thing that makes T3 Code special is how we enable you to work with all of the models, harnesses, and subs that you use. We've always had support for cloud

5:08code and codeex, but we've recently leveled up both by adding support for pretty much everything else. And when I say everything else, I mean it. We got codeex, claude, cursor, grock, open code, anti-gravity, muse code, and pi all built in directly. But if you have a different ACP compatible harness you want to use, we already have the whole ACP registry in here for things like AMP, things like Klein, things like Copilot, and so so many more. or you can

5:36add them manually yourself if you want. So pretty much everything is now supported. For those with multiple subs, I personally recommend setting up something like CLI proxy. But if you want to do it all within T3 Code, you can actually add multiple instances of the same provider and give them different O different home directories, different API keys and environment variables and more, which has been incredibly helpful for me as I rotate between accounts and I'm testing different early access type things. You get a lot of control for each of these things that you add into T3 Code, but that's not where the magic of T3 Code's

6:08handling of your subs really shines. I'll do a kind of silly demo to show what I mean. This is T3 Code working in the T3 Code repo itself. And I'm going to show you the thing that Julius was working on for the last 6ish months that finally shipped recently and by the time you're seeing this video is in the official stable release. I am currently using codeex on GBD61 medium ultraast with my codec sub. I have all of my other accounts and subs signed in here and I'm going to say I want you to review this codebase with four different models. Spin up a delegate for 61 soul

6:40high in codeex muse wonderful full at xigh gemini for argan in anti-gravity and sonnet 55 and claude. Tell them all to explore the codebase and find interesting things as well as some easy wins. This just created a work tree from the repo I was in in a directory specific to T3 code. So you don't have to worry about it clogging up your system. And now it spun up four sub aents all with different harnesses all within T3 code itself. So now I have GBT61 Soul Muse Wonderful Full Gemini 4 Argan and Sonnet 55. And these are all using their official harness and the subscriptions that I have for them. And

7:19I didn't do anything special to configure them. Gemini is a little annoying. You have to click a sign in button in the settings. But the rest are just using the stuff that you already have installed on your computer. So if you install cloud code or codeex and you sign into those there, now you can just

7:32use T3 code and it will work directly. The magic comes from this capability to have them all work together though. Not only can it spin up sub aents in a given thread in other harnesses and models is even more fun when you let the agents spin up and manage and talk to other threads. Right now all of this is contained within the one thread here, codebase findings and easy wins. I spun this up on the T3 codebase. So I'm going to tell it to spin up another T3 code thread in my lake bed repo with the same request. It can list the projects and launch another thread which direct

8:03links. And now I have same request here in a different thread for a different repo going off the same way. And these can talk to each other too. I can tell one thread to talk to the other. You can even copy the ID for a thread and give it to another and allow them to directly know which each other are. They can search and find each other. they can do a lot. This isn't like a prescribed workflow we're giving you. We're not telling you you have to work a certain way with this. The orchestration layer Julius's built here is a very flexible MCP that allows your agents to do all

8:34sorts of cool stuff. You can even expose this connection to another tool like Grockbot and manage your T3 code threads entirely externally. I know somebody who doesn't use T3 code directly at all, but has done like 10,000 threads in it because they do it all through their instant connection over iMessage, which manages T3 code directly instead. And this all just works out of the box. If you go to the T3 Code site, download the app, launch it on your computer, if you already have these harnesses set up and installed, it will import your existing threads and projects and give you this capability with a UI that, in my

9:06opinion, is a lot less annoying. It's also much faster, which I want to put in my list of three things, but I won't. There is so much more to say about this orchestration system. We've called it orchestrator v2 for a while, but I want to go through so many other things. So, let's go back to the top three big parts, and then we'll dive into all the fun details. The first thing was working with all of your harnesses, models, and subs. Again, you have to bring your own subscriptions. You cannot pay us for access to AI. If you thought you could, you're misremembering that was T3 Chat,

9:36not T3 code. D3 chat is a closed source alternative to chat GBT. D3 code is an open- source way to work with your AI agents, models, harnesses, whatever. I'm going to kind of blast through these top level big things, but make sure you stick around because the layered real fun stuff underneath is where most people find the magic. This is more to show you the the big things we are thinking about when we build and then all these other features are going to make a lot more sense. The next piece is the one I push the hardest for that has

10:04had the most impact on how I build. Worldclass remote control. What do I mean by this? You've already possibly used the remote control feature in cloud code. All you have to do is type / remote and now you can manage that thread from your phone. Isn't that all you ever need? I'm not spinning up my threads via SSH anymore. I'm not spinning up my threads in terminals anymore. I'm spinning up my threads by clicking the new thread button or the hotkey for it. And then I choose where I want them to run. T3 Code's architecture allows this to be trivial because unlike all of these other apps that are firing

10:37off commands directly over your terminals or are trying to connect over weird electron like RCP calls, we don't do any of that. Everything goes through our websocket server. If you don't already have T3 code installed and running, the easiest way to see this is go to your terminal and write npxt3 serve. And now you have a hosted local T3 code instance that doesn't require an app to be installed. If you run this command on a remote server, you can do the same there. The way this works is the T3 serve command is exposing the capabilities of T3 code over websocket and it's hosting a really minimal web

11:11app that you can go to and see. This also works incredibly well with things like Tailscale, which allows you to run one command on the machines you want to have running T3 code and then you can connect to those over Tailscale and control them fully remotely. But this just lets you go to this one web page. You don't want to manage all these different pages for all these different things, right? Well, thankfully, you don't have to because T3 Code lets you pick which of your machines every thread runs on when you spin up the thread. So, I have my main laptop called mainbook, but I also have two cloud VPSs, Alvin

11:41and Cup 2. I have my local server that also runs all my inference and stuff, BB1. Then, I have uh two mini PCs that just like random Linux boxes running T3S. And then I have a couple other Macs for when I need to test iOS stuff, Mac OS stuff, desktop apps with computer use, those types of things. So I pick which of the machines I want the threads to run on. If you don't want to have to manage all of this in your head, figuring out which machine which thing should go to or which one has enough load available, Maria built a feature for this load balancing where we'll

12:10actually look to see how much throughput is available on a given machine like how screwed is the CPU in memory and it will route your requests to whatever machine has the most available. Along the same lines as our device hub capabilities that let you run your iOS and Android simulators inside of T3 Code, you can even dedicate certain machines. Like for me, I would dedicate my Macs to this so that the other computers in the fleet can SSH in to run a dev server for the iOS app on a Mac when the rest of the work is happening on a Linux box. All just built in. Julius made all of these

12:45things because he hated working with Xcode and wanted it to be easier to iterate on the T3 code mobile app, which is also worth noting. We have the best mobile app for working on your AI code tasks remotely. If you've tried to fight the Cloud Code or Codeex mobile apps to work on your stuff remote, you know how bad they are. They just do not work for this. It is hilarious that they pretend they do. They are so far from good. It's insane. Our mobile app is incredibly functional and it's also open source. So if you do notice things that are wrong, we highly recommend you throw Opus at it

13:17and file a pull request. And if all of this SSH and tail scale stuff's intimidating to you, I get it. It's not great. I hate having to run a VPN on my phone all the time. That's why we have T3 Connect, which lets you connect three devices for free through our tunneling system instead. Combining T3 Connect with some random old computer running Linux is genuinely magical. This is a Linux box on my network that I've been using for doing random Rust builds. And I'm going to run npxt3 at nightly because I'm still in the nightly, but you don't need to do this once this is out. npxt3 connect. It grabs the package

13:49from npm that runs a migration. And then it gives me this link to add it to my account. It will blur when I do this. And now T3 Connect is set up on this computer, which means I can quit from SSH, clear, never think about it again, and manage the whole machine all from my desktop app or my phone app. It is genuinely the most pleasant way I've ever found to manage all of my AI coding stuff. And I basically never run threads on my own computer anymore. I just run them remotely on these other boxes. And

14:22T3 Code has made that trivial to do. There are lots of other cool things with the way we handle T3 connect and remote that I'll talk about in a bit. Make sure I mention web hooks later. I promise you want to hear that part. It's one of the coolest things that Julius added recently, but I want to do the one other big thing before we dive into all the fun details. The simplest way I can frame this point is that we're built to get work done. The best way I could think to explain this is to show the

14:46current state of the chat GPT app. What's going on here? What work is recent? What needs my attention? What doesn't? What is from where? Where are my recent threads? What should I be prioritizing? It's impossible to know what to do here. And what I found here is that I was getting lost a lot. I didn't know what threads were done. I didn't know what threads needed my attention. I didn't know where my threads for my important projects went because when I added a new project that was temporary, it would appear at the top and everything else would get pushed down to the bottom. I don't know how to

15:17find anything in the chatbt app anymore. I've basically given up and just use search. Funny enough, they were so inspired by our way of doing this, they added a little button here, the activity sidebar, which is trying to copy a bit more of what we have done. And I'll be real, it has not succeeded. I was so hopeful for that feature. It is just it didn't work out. So, I spent way too much time thinking about this and built a very different sidebar. And I'm going to be super honest. I didn't think people were going to like this. Turns out they did. I put it under a bunch of

15:50hidden flags. I didn't want people to get forced into this because I didn't think they would like it. It's a bit of you'll have some growing pains when you make the move and I'm sorry in advance for it, but when you learn to think about your threads the way that I am pushing here, you'll understand why. Generally speaking, we have kept T3 code almost like absurdly unopinionated. The sidebar is one of the few places that we've done opinions. We are going to make it more customizable in the future, but for now sidebar is an inbox. Every time you open a thread, the goal should be to close it. Close can be you merge

16:24the change. Close can be you answered the question. Close can be the thing isn't needed anymore. But your goal every time you open a thread should be to get rid of it. And we have a lot of things built in to help. The easiest, simplest one is this idea of settling. You don't archive threads, you settle them. It feels much less permanent. All your settled threads appear at the bottom here and you just click the little check box and now it's settled and you move to the next one. If that was a mistake, you could command Z and get it right back. This is the one I

16:51spun up as a demo showing off all of the sub agent stuff. I don't really care what this said, so I'm just going to settle it and see what's next. This one is restore working thread drag references. If I wasn't filming right now, I'd probably go and see if we can finally get these PRs landed. But I am busy right now. I'm kind of filming a video, so I'm going to snooze this until tomorrow. Now, this thread is knocked into our snooze section. It will come

17:16back after a certain amount of time. This one is completed, and it was an audit of a production release I was doing on my cloud. It said, "Release is done in stable. Nothing broke." So, I will settle that one. This one has a question for me. So, it popped back up to the top with this little input indicator to let you know there's a thing that you have to answer. Where I see the 30 plus hour estimates, I don't actually know. I don't think this box has access to it. So, uh, I don't really care. That's just for a demo for another

17:41video. So, I'm going to archive that. This one is a new scheduled task that I just spun up, which, uh, by the way, we now actually support task scheduling in T3 Code. So, I'm going to tell it this looks good. Set it with a schedule. I do really like that Claude understands that compaction is a thing and it might lose

18:00instructions if it doesn't encode them. The fact the models are aware of compaction is very good. Obviously, we have steering, too. So, when the model's annoying you, you can send a message to do a follow-up. We also have queuing. So, cued message, and you can have a hotkey to swap between the two. I don't like queuing, so I set my default to be steer, but you can change that in settings. And here we now have a scheduled task. So, this thread will now run daily at 900 p.m., which means I can archive it. And now, this thread will reappear at 9:00 p.m. when it has something to do and run. Here's a thread

18:33that I snoozed earlier that I don't really care about. So, I'm going to snooze this one to be after stream. One other thing you might see here is that I have threads going, but there's nothing visible here. That's cuz we have a working section on the bottom. This is optional. I currently have it as a beta feature. The working section, this hides

18:51threads while they are actively working. I thought this was going to be another one of those changes people hated. I can't live without it now, and I've seen a lot of other people doing the same. I did not think this would be as popular as it is, but it seems like you'll love it. It's hit or miss depending on what you like and what you are. I do have a couple workarounds if you want to use it, but also keep eyes on things. If you pin a thread, it no longer will hide in that section even if it's working or waiting. So, you have ways to work

19:15around the stuff. If you're working on a lot of different things at the same time, or you have like three different things in your mind that you want to go fire, like let's say I want to uh index all my open PRs for me and help me prioritize them. Let's say I have something else I want to do right after. Instead of sending and then hitting the hotkey for new thread and starting there, I can command enter and it fires in the background and then I can do another thing and fire that the same way. And now all these threads appear under the working section. This one just

19:46finished. So it appeared at the top and when these ones finish the same thing will happen. If you press enter it sends normal and it kicks you to the thread. Command enter will send it in the background which is very nice. What you might be noticing is that T3 code was built for getting a lot of work done in parallel and to not destroy your ADHD if you're like me. I don't want things in

20:06my face unless they need my attention. And I found that every other one of these AI building and coding tools is trying too hard to put everything in your face and make everything accessible. I don't want that. If the thing is working, get it out of my face. Don't bother me until it's ready. Even better, when a PR is opened by a thread, it gets linked to the thread. You've probably already seen that in some of the threads earlier where these have this one has five PRs linked to it. This one has two, this one has one. You can click it and open it in the inapp PR

20:34viewer, which is super super helpful. Or commandclick if you want to open in your browser of choice. When this PR merges, the thread gets auto settled. Why do you need the thread anymore if the code is merged? Get it out of my face. You should only have a thread visible if there's a reason for it to be there. if

20:51it is working or the code merged. Obviously, that setting is configurable and it's even configurable on a per thread basis. Like if I don't want this one to settle itself, I can disable autosettle for this particular thread. This all works even better on mobile, by the way, because we have the swipe gestures for archive and snooze, so you can blast through the list of threads and figure out what you want to do with them and how you want to do it. All of this is built to try and keep your workplace clean and restore a sense of what I call spatial awareness. When I was working in traditional code bases by

21:22hand, the oldfashioned way, I would build up a good intuition of where things were and where to go to find them. T3 code was trying to bring that feeling back of like knowing where things are so that you aren't constantly overwhelmed with in your face. All of these flows are built to make it easier to not get overwhelmed as you are doing more and more in parallel. Would you believe me if I said we were just getting started? think it's fair to say there's a little bit more that is worth bringing up and talking about. I'm going to do my best to blast through all of

21:58it. T3 Code is truly open source. Everything we're working on is public in the repo. Even our infrastructure backends, T3 Connect and more. Everything you would ever want to see in T3 Code is here. You can fork it and run it yourself. You can even connect over T3 Connect if you're running a fork. The remote connections are great for running prompts in T3 Code, but they're good for a hell of a lot more, like your

22:16terminal. This isn't running over SSH. This is running over T3 Connect. And we have found it to be way more stable than SSH, especially on mobile. I'm running BTOP in T3 Code over T3 Connect inside of T3 Code itself. We also track full diagnostics inside of the dashboard for all of the different machines you have connected. So I can see how much compute all my boxes are using, which is the data we use to route things to the right place. Your agents aren't limited to rendering markdown in the UI. They can make full HTML visualizations with interactive JavaScript that you can actually use in the app, making charts,

22:48diagrams, and so much more. I use this pretty much every day, at least 10 times a day. Tired of dealing with poll requests in GitHub? You'll never have to again because you can do it directly in T3 Code by going to the little tab at the bottom and looking at all of your PRs, managing them, merging them, and more from here. Julius even got this all working in the mobile app, which is way better than the GitHub app, which I

23:08actually thought was pretty good. Concerned about where your tokens are going? Don't be. T3 Code has a usage tab that shows you where all of your usage is going. Even if you're not using T3 Code yet, if the environment is connected to T3 Code, we can read the usage from it. So even though a lot of my cloud spend here is with a Cloud Code terminal run in a loop, it all shows up here the same. And if you're one of those people that's thinking about limits and isn't using a proxy yet, all of that can be shown inside of T3 Code as well. I just merged a change to DDUP

23:33accounts cuz people were having issues with that. I don't think I got that in this build yet, but you should very clearly be able to see what each of your accounts has left. I don't sign into my accounts directly, so I can't show this next feature. But do know that if you're hitting your limits, you can schedule a follow-up message to auto happen when your limits are available again. So if you're rate limited on the 5 hour and you can't send a prompt for 3 hours, you can send it now and it will fire as soon as your limits are available again. You might have noticed as I've been flying

23:57around the UI, but T3 code is disgustingly fast. I have put way too much time into minimizing the amount of data necessary to transfer to do things. And hopping between threads uses a hundred times less data than Codeex does, possibly more at this point because I know they've gotten a bit bloated and we've gotten a bit trimmer. That means we work great on crappy mobile connections. Trying to send a prompt on 3G. I promise T3 Code will be the best experience you've had doing that. I code on airplane Wi-Fi and not Starlink, like the crappy airplane Wi-Fi, a lot. And I don't even notice until I command tab to the browser and

24:28everything is suddenly slow again. This one is small, but I'm amazed nobody else has this. You can shiftclick to select multiple threads at a time and snooze or settle all of them at once or regenerate the titles if the titles are no longer accurate for what those threads are doing. So I can click regenerate and these will regenerate with different titles. You can even pick which harness, model, and effort level is being used for that text generation for things like commit messages, titles for threads, PR descriptions, and more. Threads will auto settle if you haven't touched them for an amount of time. And you can turn

24:58this on and off and configure how long until they settle. The default is 3 days, which I think is fair. I bumped it to seven because I want my inbox to get full, so I'm forced to deal with it. If you don't want to install the app, don't worry, you don't have to. Go to app.t3.codes in your browser of choice, and you can run T3 Code in the browser and connect to your Npxt T3 instances that you're running on other machines, making it trivial to access all your boxes. This was the main way I used T3 Code for a while, funny enough. Typing a prompt and have a different thing you

25:24want to write up, just run the hotkey again, and now you have another thing. And these drafts get stored at the top here. And they're very easy to delete. And if you want to send this somewhere else, like you pick the wrong project or thread, command S to stash it. Go somewhere else. Command S to apply the stashed thing. You can see them all here. You can even stash images. Not only does T3 Code have best-in-class mobile emulator support built in. It also works remotely. This isn't running on my computer. This is from another computer on my network right now over T3 Connect. It's not even a direct

25:55connection. And I have a full iOS experience here running remotely. You can even do this on mobile. You can interact with a mobile iPhone on the iPhone app. And by the way, this is a full 3D experience, too. Jesus, I didn't even know this. You can two-finger to look at the full 3D device. Remember, still fully remotely. Do you know? God, this is so cool. Especially when you're using something like the iPhone Duo simulator, which I know Julius is working hard to make sure we have day one iPhone Duo support that is best in class. Like Xcode wishes it could. By the way, the same thing works with the

26:32built-in browser where it runs on the remote server but pipes in the results to your experience remotely so that you can have a computer that is headless running things in a browser and still be able to interact with it, sign in, and more. I wanted to show off the MCP server in my Grockbot, but if I open up my Grockbot, you'll see things that are private because I use it a lot. But T3 Codes MCP server that enables threads to interact with each other and also spin up sub agents and such also lets other

26:58apps connect to T3 code and manage it. So if you want to access and manage your T3 code instance in Grockbot with your dot in chat GPT or something else like that, you can fully control T3 code over the MCP. I'm not sure if you noticed this thread is still going and it's able to fully interact with, connect to, and manage working with T3 code like this remotely. You could even float the device over chat if you want to work with it this way instead. Another fun thing we recently added is the ability to switch between harnesses. Previously, you'd be stuck on whatever harness you started on because that's just how they

27:30work because the thread existed inside of the harness. We've built a system to allow the context to break out and go somewhere else. So I can select opus 5.5 and say see anything wrong with the app send. And now the context has been handed over to Opus which will take over accordingly. And it still has the same connection to this iOS device and it's able to work accordingly. For better or worse, we support far too many OSs. I hate hate hate Windows ARM so much. It has massively increased our build times and fails all of the time. But we support it anyways because whenever I deprecate one

28:09of these things, people get mad at us. Almost half of the events sent in T3 code are coming from Linux users, which is insane. Linux is very, very well supported. It's not just like a thing we threw together and pretend that we support. It is core to T3 code to the point where it's actually built into Omari now. Yes, for better or worse, T3 code is in Omari directly. Personally, I run most of my agents on Linux, but I don't like using Linux on the computer that I am touching and typing on. So I am able to control all those Linux boxes from T3 code on my Mac. Want a thread to

28:40have context on another thread? Drag it in. Now this thread knows about this other thread and you can say what is this thread and it now knows and can interact with it. You can even tell it to talk to the other thread after you've done that. And if you want to comment on something that was said here or something in the browser preview, you can select it, click site and say I don't agree, save, and then send all of these annotations along with your next prompt. And how can I forget about themes? I know not everybody likes our colors, but if you liked T3 Chat's colors, you can switch to that. If you

29:10want a different theme for light mode and dark mode, you can do that, too. You want to import a theme from VS Code, we support it. You want to make your own, we support it. Want to make your chat width disgustingly wide, feel free to. I'll judge you, but you can do it if you really want. We have unnecessarily configurable key bindings that you can set across your machine or across all of your machines in your environment. You might have noticed every settings page lets you pick if this applies for one project and if this applies for specific environments. So you don't have to worry about a setting only being on one thing

29:41and not another. You have full control over where everything goes. You can even see which settings are on which machines. You can control when, how, and where work trees are spawned. By default, I highly recommend that you spawn work trees because it just makes managing everything way easier when you're working in parallel, but that's your choice. You can also start threads

29:57with no project or with a new project. You give a new project a name. You pick which machine it goes on. And you also pick whether or not it gets autosynced up to GitHub, making it very easy to spin up a new project on whatever machine with whatever idea you have. This alone has made me spin up so many more projects that I never would have before. This one hurts me a bit, but we support pretty much every Git provider

30:20you would reasonably want to use. whether it's Azure DevOps, GitLab, Bitbucket, Forge Joe, Git Cafe, and of course the loved and not at all frustrating GitHub. I also put an actually disgusting amount of time into our PR monitoring flows that are built into T3 code directly now so that instead of having your agent reinvent babysitting every time you ask for it, you can use a reliable method that I built that uses as little as a hundth as much of your GitHub API rate limits as whatever the your agents are throwing together are. This is a small one, but it's one I really like. You can set up multiple routes to connect to a

30:57given machine. So you can have it go over tail scale, T3 connect, over LAN, or whatever other methods you would like to connect with. And you have a very granular permission system to control what permissions are and aren't allowed over these remote connections. You can also one-click the machine and click copy MCP URL to expose that MCP server to other machines. Since you can connect over T3 Connect, you can also expose a thread over a web hook. This allows you to have threads that trigger on certain web hook actions like when a PR is opened or when something occurs in Stripe. You get a URL that forwards

31:29events to your environment or you can opt in to have them held and replayed for up to 24 hours for added durability if your servers down. You can even bring your own public URL. This lets you fire a curl to trigger a thread remotely. It's so cool. And of course, thank you all for reminding me folder icons. When you add a project to T3 Code, if it has a favicon, that will become the icon for that project as long as it's in a normal place. And if it's not, you can rightclick project settings and then choose the icon depending on where it is

31:56or even generate one with our little UI. Theo from the future here. There's a few features I forgot and a few more that we've added since I originally filmed that I wanted to make sure I covered here. The first one is very useful right now because I have a lot of machines and I want to update all of them. So the update button in T3 Code not only will update the app on your computer, it will also update the providers. So if you're using Claude, Codex, or whatever else, those can get updated as well as the version of the app on all of those machines as well. With one click, I'm

32:22updating Claude and Codeex and T3 Code across nine different computers. This is the one I was so mad I forgot that I decided to record the extra. We have an open and VS Code button. This obviously works with other editors, but you might be confused. Why am I so excited about an open VS Code button? It's because it

32:38works. Like obviously it works, right? I'm just opening a project up. Did you notice that this project isn't on my computer though? This is on a different machine, BBOX. We have T3 Code set up so that when you open something in VS Code, as long as your client has SSH access to the other machine, then we'll automatically open VS Code in SSH mode connecting to the machine remotely. So, I can edit all of the code however I like to. On the rare occasions I am looking at the code, I'll be real, I mostly use this for environment variables and occasionally markdown changes. But it's so nice being able to

33:08one-click open over SSH to whichever box the code lives on. On the topic of those of us who use multiple machines, I recently added the ability to port projects from one machine to another. So I have TS Rust on a couple of my boxes, but if I wanted to go spin this up on Leftbook or BB-1, I can just click here and it will set up and clone the whole repo on whichever box I click. You could even choose to copy environment variable files and other git ignored content if it exists in the repo when you transfer it. And in just a few moments, the repo

33:36is fully cloned and managed on another machine. And now I can send prompts to whichever box I want. And if anything does go wrong and the app isn't working, hop to your terminal and type in npxt3 triage. You pick between cloud code and codeex and it will pass all the necessary context for you to debug your problems on your machine yourself and even cut an issue if it turns out the thing you ran into is universal and should be fixed in T3 Code. We closely monitor the issue. So if you do find something this way, make sure you report it. We'll get it fixed as soon as we

34:04can. Since T3 Code integrates with so many different things, there will always be some niche edge cases and bugs, not only can T3 Triage identify those and report them to the team, it can also unblock you most of the time locally to get things fixed on your machine for your specific setup. From what I've seen, over 80% of the issues people have can be successfully resolved just by running this one command and giving the model a bit of context. I think that's everything I forgot. You can go back to

34:28whatever Theo was just yapping about. People love the web app. They love HTML and chat. They love file attachments. Oh, file attaches is another really annoying one that I put a lot of time into. When you attach an image in T3 code, it does pass up the image the way everything else does, but it also puts the file somewhere on the system that your agents can access. So, if you paste an image into your chat and say add this to the PR, it can. None of the other tools can do that. They can see the content of the image, but they don't give the actual bites to the model so

34:58that it can do things with them. We do. That means you can attach anything. People love the PR viewer. People are excited to use the MCP and web hooks. D3 app can render videos and photos over these remote connections, which is super useful if you wanted to add and attach a video of a fix that it made into a poll request. The built-in monitoring people really like the GitHub CLI can't upload images. It kind of can now, but your agents don't know about it. You have to be on the latest version. So, I have a file upload skill and my agents know about that and then they can use that to

35:26embed. People really like the custom rendering stuff that we have. Oh, snapshots. I forgot about snapshots. How did I forget about snapshots? Because I don't have it on. Snapshots lets you press both shift keys to add some context to T3 Code. So, if I'm in the browser and I want T3 Code to have this info, I can press both shifts. It grabs it and then opens a thread in T3 Code

35:45with the things that I just snapshotted. It's not just the content of the app that you were in. It's also a bunch of the other accessibility data that's available on your system. Currently, our custom font support is better than everything else. That's cool to see. I'm just using the defaults. Since it's open source, your agents can customize the hell out of it. We're working on some real fun customization systems in the future where you can modify your direct app on your machine instead of having to clone it. And if you somehow manage to break things really badly, you can run npxt3 triage in your terminal. It will

36:11kick off all the context needed to cloud code or codeex in order to diagnose whatever the hell went wrong. Not only will this let you fix whatever's broken locally, it also makes it easy to file a detailed issue with all of the necessary logs so that we can actually fix the problem. The live activities in the mobile app are incredible. I don't have notifications on for my mobile apps. I only want to look at T3 Code when I'm using T3 Code. But if you are the type of person that wants to know how your agents are doing, you can have that just in your notifications on your iPhone. I

36:38just learned that there's an option in the PR tab to sort by merge readiness. That's pretty awesome. I don't know how it's measuring that, but that is super helpful. We've been being too nice to T3 Code. Let me go through the negatives fast. I actually wrote them down. Uh, you have to bring your own compute. So, we don't have some cloud provider that we're overcharging for. You need to have something that T3 Code runs on. You can spin up in your own cloud if you want. I don't care, but you need to do that somehow. T3 Connect is currently limited to three devices, which means you need

37:04other ways to connect to your remote environments if you have more than three. So, if you want to have 17 computers, you better know how tail scale works. In the near future, we are going to have something to let you have more than three devices. We just don't want to bankrupt ourselves on an open source project. Computer use isn't built in yet. You can absolutely use the computer use if the harness is providing it. So things like codecs support this by default. You can actually take the codeex apps MCP for computer use and give it to Claude and then Claude will work too, but we'll have our own

37:31built-in in the near future. This is probably the worst part by far. It's made by Theo, which sucks. Probably the biggest issue with T3 code, honestly. Thankfully, it's open source, so you can fork it if that bothers you enough. It will ruin your sleep schedule. I know this has happened to a lot of people. Uh chat, let me know. Has T3 Code affected your sleep? ones in chat if T3 code has affected your ability to go to bed on

37:52time. Yeah.

37:58Yeah. So, if you want to go to bed at regular times and not have your addiction to vibe coding keep you up all night, maybe don't try code. I talked about it with my therapist. Yeah, you guys don't have phones in bed. Yeah, but then I stay up all night prompting from my phone in bed instead with the T3 Code mobile app. I had to have a conversation with my ex about my inability to get off my phone in bed

38:25when I was prompting T3 Code all night. I'll just say there's a reason they're an X now. And of course, you will burn more tokens. Not because T3 code is super token inefficient or something. I've this heavily. T3 code will use your usage just as fast as you would get it burned through claude or codeex directly, but you'll be firing off so many more threads that you'll feel your usage burning faster because you're going to be doing so much more. I think I've said all I have to. If there's anything T3 code doesn't do that you really think it should, cut an issue or join us in the T3 Code discussions on

38:58GitHub. You might be surprised how many others want it and if you can motivate me enough, we might end up adding it too. Generally speaking, T3 code is built for developers to make them more effective at development. And it feels like the rest of the world is going the other way. Where chat GPT is trying to shove dots and spaces and plugins in your face. Cloud code's trying to be for

39:22work, not for code, which is obnoxious. Everybody else is forcing code into a tab in their other apps. T3 Code is built for code. T3 Code was built by engineers for engineers. and we are not compromising on that. So if you miss when these tools were built for you as a developer, T3 Code was built for you as a developer. And of course, the most important thing, T3 Code was built with a team of worldclass engineers that totally aren't Twitter posters

39:51full-time. Definitely not posters. Definitely just engineers. I hate this thing so much. Bring it up back. If you download T3 Code today from the site, you are going to get the stable version. By the time this video is out, the stable version will have all the cool things I just showed. But if you want to be on the bleeding edge, get all the cool things right when they're ready, and you're okay with things changing as we iterate and potentially breaking for really short windows when we make these types of changes, click that little nightly button on the download page, and you'll have an experience that at the

40:21absolute least will get updated a hell of a lot faster. I can't believe it took me this long to do a dedicated T3 code video, but I figured it was finally time. I'm mostly doing this because my team has been working so incredibly hard and I feel bad not showcasing their awesome work. So, I'm going to wrap this up with a shout out to all of the awesome people who have been contributing heavily to T3 Code. Thank you to Julius for pushing this project as far as it has gone today. Thank you to Maria for really helping take charge on the insane onslaught of PRs that

40:47we've been dealing with. Thank you to Shiv for coming in clutch and helping clean up a lot of stuff throughout. Thank you to WOW for building a ton of the integrations around PI and Orchestrator V2. Thank you to Bal for building the original version of the PR viewer. It is incredible. It is so cool. And of course, thank you to all of the other contributors. I'm forgetting so many of y'all here. There are 516 of you all now. To be fair, so I hope you can understand how hard it is to keep track of it all. And of course, thank you to all 27,000 people who have given us a

41:18star on GitHub. And if you haven't yet, we would appreciate it. I think I've said all I have to on this one. It's honestly kind of surreal to see how positive the reaction to all of this has been. I just wanted to make sure that the future tools we build with wouldn't be super expensive private closed source things. And I'm thankful that the future for open source is bright and that so many people are jumping to support the

41:42chaotic things that we're building. Thank you all for coming on this wild ride with me. I can't wait to see what we do next. And until next time, peace nerds.