T3 Code has potential... (Better than Codex?)

BBetter Stack
Computing/SoftwareSmall Business/StartupsInternet Technology

Transcript

00:00:00Theo just dropped T3 code and I am pretty excited for this one. I mean, what's not to
00:00:04love about it? It's a fully open source agent manager that is completely free. It just uses
00:00:08your existing subscriptions. And I know terminal UIs are all of the hype at the moment, but
00:00:13I'm sorry for multiple agent workflows across multiple projects. A GUI is just a better experience
00:00:18for me and that's why I've been loving the Codex app recently, but I have been annoyed
00:00:23by its performance. It absolutely loves RAM and also the fact that you were tied down
00:00:27to OpenAI's models. So let's see what T3 code has to offer and if it solves some of these
00:00:32pain points.
00:00:38So this is T3 code and if you've used the Codex app before, this UI is going to feel
00:00:41pretty familiar. On the left we have our projects, then inside of them we have the individual
00:00:45threads that we have open. But one of the first things I want to clarify is what this app actually
00:00:51is. This is not competing with Claude code or Codex as a coding agent. This is simply
00:00:55a GUI on top of those tools. So all of the code that you see in this thread and all of
00:01:00the responses are coming from Codex behind the scenes and therefore it's using my OpenAI
00:01:05subscription. That is why T3 code is completely free. If we go down to the model selector,
00:01:09you can see I can choose between the available models on OpenAI and you can see they want
00:01:13support for Claude code, cursor, open code and gemini in the future. In fact, I've actually
00:01:18seen the Claude code support is ready. They're literally just waiting for clarification from
00:01:21Anthropic to see if they can use Claude code subscriptions this way. For me, this is going
00:01:25to be one of the biggest advantages of T3 code because while I do like OpenAI models for coding,
00:01:30some of the tasks that just were sat like UI design. So I do have to switch apps from time
00:01:34to time and open up core code in the terminal while I'm working with the Codex app. If they
00:01:39get this integrated, I won't have to, I can manage it all from one place. Plus I also tend
00:01:43to find that the models creators make the best harnesses for them. AKA OpenAI models work
00:01:47best in Codex and Anthropic models work best in Claude code. So this way we still get the
00:01:52same quality. This also means that T3 codes focus is on the user experience of managing
00:01:56and using these agents. So if I jump into a project I was working on earlier, this is actually
00:02:00my personal blog, which I haven't upgraded in years. So I want to change the way that
00:02:04markdown works in it. Now down here we have our reasoning selector alongside our model
00:02:08selector and also fast mode on or off. If you're using the OpenAI models, then we can choose
00:02:12between chat and plan mode as well as full access and supervised for those tool calls.
00:02:16I'm actually going to leave this on plan mode and over here we can choose between local and
00:02:19a new work tree. I really like get work trees for working with agents. So I'm gonna leave
00:02:23this on new work tree and also leave this on the main branch in the selector over here.
00:02:28So if I send off this prompt, Codex is starting to look around my repo and build out a plan
00:02:31for me and we can see the tool calls being streamed in here as well. I also set off two
00:02:35other agents in these projects here to get on with some coding work for me. So we can
00:02:38see that with a nice working badge. While the plan is still building, I want to point out
00:02:41that I really liked that this automatically picked up my favicon for this project. It's
00:02:45just a small touch that I think is pretty cool. After a few minutes, the agent is asking
00:02:48for some user input. So we have a question here. I'm gonna go ahead and select that answer.
00:02:52And after a few more minutes, we can see it's now done with its plan and we get to see that
00:02:56in the chat here. We can also download this as markdown or save it to the workspace and
00:03:00expand it if you want to read the entire plan. I'm just gonna go ahead and implement this.
00:03:04And while that does my work for me, I want to talk about the other selling point of T3
00:03:07code and that's its performance. If you've ever used the Codex app, you may have noticed
00:03:11if you have multiple agents running, it starts to get a little bit laggy and I've even had
00:03:14to restart my Codex app because it just starts to lag out completely. I've actually been using
00:03:18T3 code with multiple agents for a few hours now and I haven't experienced that once. You
00:03:22can even feel the attention to performance because if you click around in these threads
00:03:25here, you'll notice they all load immediately, even if they're super long threads. Back to
00:03:29my project though, it looks like it's done with the code changes. So we go up to this
00:03:32menu here. We can click on commit and obviously see an overview of the files that have changed.
00:03:36And if we leave this message blank, it's going to auto generate one for us. We check out the
00:03:40menu here, you can see we also have push and create PR. If you click this button on its
00:03:44own, it's going to generate the commit message for you, push it, and then open up a new PR
00:03:48straight away. Once it's created it, you can see we get an icon in the sidebar over here
00:03:52telling us this has an active PR for this thread. And we can also just click on this button to
00:03:56be taken straight to the PR itself. It might seem like quite a minor feature, but I just
00:03:59think it helps you ship faster by just giving you that button to straight away push the code
00:04:03up, open up a PR and you can move on with your project. You'll also notice that this icon
00:04:07will update based on the status of your PR. So if it's rejected, this will change red.
00:04:11And if it's merged, you can see it's changed to purple. As for the other buttons that we
00:04:14have above a thread, if we click on this one, we can see a diff view for each of the individual
00:04:17turns that the agent took or all of them combined. I think this renders in very nicely and we
00:04:22can choose between a split or a unified view. Then we also have a simple button to open up
00:04:26the code base in cursor or finder. Then we have the quick action button. If you haven't
00:04:30seen this before, this essentially just allows you to set up a button that's going to go ahead
00:04:33and run a command in the project for you. You also choose if you want to run this automatically
00:04:37on a work tree creation or if you want to create a key bind for it. So with that one set up,
00:04:41we now have a button. If we click on it, it runs one install within the project that we're
00:04:44in. And it also brought up the integrated terminal that's going to show you next anyways.
00:04:48Not much to say here besides it's a normal terminal just means you don't have to leave
00:04:51this app very much. The other feature that I really like that's quite a small quality
00:04:54of life improvement, at least for me, is when you click on add project, instead of browsing
00:04:58your finder files, you can actually just paste in a path to the project. I quite like this
00:05:02because often I scaffold my projects out in the terminal first. Now I can simply copy the
00:05:05working directory and paste it in to open it up. It does kind of make me wish though that
00:05:09there was just a terminal command I could use that would do the exact same thing. So
00:05:12that's definitely the first item on my wishlist. So now let's talk about a few more features
00:05:16that I think are missing from T3 code. I'll start off with four minor quality of life improvements.
00:05:21The first one is I want to be able to double click on a thread title to rename it instead
00:05:24of going into the menu. The second one is in this menu over here. I want to see an option
00:05:28to open it up in my terminal. The third one is I want to be able to collapse the sidebar.
00:05:32It doesn't seem like you can do that at the moment, which seems like a bit of an oversight.
00:05:35So I'm sure they'll add that soon. And the fourth one is I want to be able to see the
00:05:39running terminals at the moment. Localhost 3000 is running in this session, but I see
00:05:43no indication that there's actually an active terminal. The good news is though, since this
00:05:46is open source, I can actually just clone the repo and ask codecs to add these features for
00:05:50me. So after a bit of vibe coding, I now have double clicked to rename so I can change this
00:05:54to subscribe. Something you should definitely do. I can now click on this menu to open it
00:05:57up in a terminal. I can see the running terminals over in this session here, and I can also do
00:06:01command B to close the sidebar. It does seem like I have a bit of a formatting issue though.
00:06:05Now besides those poorly vibe coded features I just added, there are still some important
00:06:08ones that I think are missing, and the first big one is better skill support. I'd love to
00:06:12see a way to be able to tag your skills in the prompt like you can in codecs, and also
00:06:16a view to see all of the skills that you have installed, and the same thing goes for MCP
00:06:20servers. I'd also love to be able to run a headless version of this on a dev machine in
00:06:23the cloud so I can set up and manage remote agents through it. I'm sure these will be added
00:06:27soon though, as the repo already has a ton of feature requests, and Theo and his team,
00:06:30especially Julius, ship very fast. But let me know what you think of T3 code in the comments
00:06:35down below, why don't you subscribe, and as always see you in the next one.

Key Takeaway

T3 Code provides a high-performance, open-source graphical interface that centralizes multiple AI coding agents and subscriptions into a single, efficient workspace.

Highlights

T3 Code is a fully open-source GUI for managing coding agents like Codex and Claude Code.

The tool is completely free as it leverages the user's existing API subscriptions for models.

Key performance advantages include better RAM management and faster thread loading compared to the Codex app.

Integrated features like automatic PR creation, git worktree support, and a built-in terminal streamline the development workflow.

The open-source nature allowed the creator to 'vibe code' missing features like sidebar toggling and thread renaming immediately.

Timeline

Introduction to T3 Code and the Shift to GUI

The speaker introduces T3 Code, a new open-source agent manager released by Theo that focuses on providing a superior graphical user interface. While terminal UIs are currently popular, the speaker argues that a GUI is more effective for managing multiple agent workflows across various projects simultaneously. He highlights specific frustrations with the existing Codex app, particularly its high RAM consumption and its limitation to only OpenAI models. T3 Code is positioned as a solution to these performance bottlenecks while offering more flexibility in model selection. This section sets the stage for a tool that prioritizes user experience and resource efficiency in AI-assisted coding.

Architecture and Model Support

The speaker clarifies that T3 Code is not a standalone coding agent but rather a sophisticated GUI wrapper sitting on top of tools like Codex and Claude Code. Because it uses the user's own subscriptions, the app itself is free and doesn't require a separate monthly fee. There is a strong emphasis on future-proofing, with planned support for Cursor, Gemini, and Anthropic's Claude Code once licensing is clarified. The speaker notes that using native 'harnesses' ensures the best quality, such as using Anthropic models specifically through Claude Code for UI tasks. This architecture allows developers to manage all their specialized AI tools from a single, unified interface without switching apps.

Live Demo: Project Management and Workflows

Using his personal blog as a test case, the speaker demonstrates how to set up a coding task using the reasoning and model selectors. He explains the benefits of using 'git worktrees' for agents, which allows the AI to work in an isolated environment without cluttering the main local branch. The UI provides real-time streaming of tool calls and even automatically pulls in project favicons for a polished visual experience. During the demo, the agent builds a plan, asks for user clarification, and presents a markdown summary of the proposed changes. This workflow highlights the interactive nature of T3 Code, where the user remains in control of the implementation phase.

Performance Gains and GitHub Integration

A major selling point discussed here is the app's snappy performance, even when running multiple agents across long conversation threads. The speaker contrasts this with the Codex app, which often lags or requires restarts during heavy multi-agent sessions. Beyond speed, T3 Code includes 'ship faster' features like automated commit message generation and one-click Pull Request (PR) creation. The sidebar icons dynamically update to show the status of these PRs, changing colors to indicate if a request was merged or rejected. These integrations are designed to reduce the friction between writing code with AI and actually deploying it to a repository.

Development Tools and Quality of Life Features

The speaker explores the technical utility of the app, including a built-in diff viewer that supports both split and unified views for reviewing agent changes. He demonstrates 'quick actions,' which allow users to map custom commands or keybinds to specific project tasks like running 'npm install.' An integrated terminal is included to ensure developers rarely have to leave the T3 Code environment to execute basic shell commands. A small but appreciated feature is the ability to paste file paths directly when adding projects, which caters to developers who scaffold projects in a separate terminal. However, the speaker notes a desire for a dedicated CLI command to open projects directly from the outside.

Open Source Customization and Future Roadmap

In the final segment, the speaker identifies missing features like sidebar collapsing and thread renaming, then proceeds to add them himself by cloning the open-source repo. This 'vibe coding' session demonstrates the power of the community-driven model behind T3 Code. He concludes by outlining a wishlist for the official roadmap, including better Skill support, Model Context Protocol (MCP) server integration, and a headless version for remote dev machines. The fast shipping pace of Theo's team is cited as a reason for optimism regarding these advanced features. The video ends with an invitation for viewers to share their thoughts on the tool's potential to disrupt the current agent landscape.

Community Posts

View all posts