Theo-brainUnofficial. Not affiliated with Theo or t3.gg.
Sign inYou
41:49
0:00

T3 Code broke me

Thursday, Oct 29 · 42 min · about 10 min to read

Theo-brain watched all 42 min

Theo tours T3 Code, an open-source app for running AI coding agents that works with your existing AI subscriptions, like Claude Code and Codex, in one place. The ideas worth taking are running agents on other machines instead of your laptop, mixing models in one task, and treating agent threads like an inbox you empty.

Skim itMost of it is a fast feature list the notes cover, but the demos of multi-model agents, the inbox sidebar and the honest downsides are worth seeing.

Tap any time to jump the video there.

Key takeaways

Theo's simplest pitch for T3 Code is an open-source copy of OpenAI's Codex app that works with all your AI subscriptions. He says that pitch gets people to try it, but it isn't why he uses it himself.

thing. The easiest way to explain it is that it's an open-source alternative to the Codex app that supports all your subs. I find that this is the best way to 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

Agent threads can start, find and message each other through a flexible MCP (a standard plug-in connection for AI tools). It is not a fixed workflow. Outside tools can even drive T3 Code: one person ran about 10,000 threads purely over iMessage.

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 built here is a very flexible MCP that allows your agents to do all sorts of cool stuff. You can even expose this

You can switch a running thread to a different model and tool mid-task. Theo handed a thread over to Opus 5.5, and it kept the context and its connection to the iOS simulator. Normally a thread is stuck inside the tool it started in.

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

The downsides he lists: you supply your own computers, since there is no hosted cloud. T3 Connect is capped at three devices, so beyond that you need Tailscale (a private-network tool). There is no built-in computer use yet.

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

Expect to burn through your AI usage faster. Theo says he audited it and T3 Code uses no more tokens than Claude or Codex directly. You simply fire off many more threads.

audited this heavily. T3 Code will use your usage just as fast as you would get it burned through Claude or Codex 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

Two skills here are new to you.

One task, several models at once

What it is
From one thread, ask the agent to spin up helper agents on different models. In the demo these were GPT-6-1, Muse, Gemini 4 and Sonnet 5-5, all exploring the same codebase for interesting things and easy wins.
Why Theo cares
Each helper runs in its official tool on the subscription he already pays for, and he configured nothing special. Gemini was the exception: it needed one sign-in click in settings.
How to try it
  1. Install Claude Code, Codex or others and sign in there; T3 Code uses those existing logins.
  2. In one thread, ask for delegates on several named models to explore the codebase.
  3. Ask them to find interesting things and easy wins.
here, and I'm going to say, "I want you to review this code base with four different models. Spin up a delegate for 6-1 soul high in Codex, Muse wonderful at x-high, Gemini 4 Argon and anti-gravity, and Sonnet 5-5 in Claude. Tell them all to explore the code base and find interesting things as well as
these are all using their official harness and the subscriptions that I have for them, and I didn't do anything special to configure them. Gemini's 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.

Run agents on other machines via T3 Connect New

What it is
Install T3 Connect on any spare computer, such as an old Linux box, with one command. Then manage that machine's agent threads from the desktop or phone app. It is free for three devices.
Why Theo cares
He calls it the most pleasant way he has found to manage his AI coding. He basically never runs threads on his own computer anymore.
How to try it
  1. Have your agent run "npx t3 connect" on a spare computer and add the link it gives to your account.
  2. When starting a new thread, pick which machine it should run on.
  3. Check on and steer those threads from the T3 Code phone app.
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 T3 code has made that trivial to do. There

Treat agent threads as an inbox

What it is
The sidebar works like an email inbox. You "settle" finished threads (a softer archive) or snooze them until later. A thread settles itself when its pull request (proposed code change) is merged.
Why Theo cares
Running many agents at once, he got lost in other apps. The rule is that every thread you open should end up closed: merged, answered or no longer needed.
How to try it
  1. Settle each thread once its change merges or its question is answered.
  2. Snooze threads you can't deal with now, such as until tomorrow.
  3. Send prompts with Command-Enter so they run in the background and stay out of view until done.
the sidebar is an inbox. Every time you open a thread, the goal should be to close it. Close can be you merge the change. Close can be you answer 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
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

npx t3 triage for when the app breaks New

What it is
A terminal command that hands Claude Code or Codex all the context needed to debug T3 Code on your own machine. It can also file an issue to the team.
Why Theo cares
T3 Code connects to so many tools that odd bugs are inevitable. He says over 80% of people's issues get fixed just by running this command.
How to try it
  1. If T3 Code misbehaves, have your agent run "npx t3 triage" in the terminal.
  2. Choose Claude Code or Codex to do the diagnosis.
  3. If the problem is not just your setup, let it file an issue.
the app isn't working, hop to your terminal and type in NPX T3 triage. You pick between Claude Code and Codex, and it will pass all the necessary context for you to debug your problems on your
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

Advice

Use worktrees by default. A worktree is a separate working copy of the code for each task, so parallel agents don't trip over each other. Theo says it makes managing parallel work far easier.

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

Don't have your agent build its own pull-request babysitting each time. Theo built a monitoring flow into T3 Code that uses as little as a hundredth of the GitHub API allowance that improvised agent versions burn.

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 hundredth as much of your GitHub API rate limits as whatever the your agents are throwing together are. This is a small

Theo sets new messages to "steer" the running agent mid-task rather than queue up for after it finishes, because he dislikes queuing. You can change this default in settings.

do a follow-up. We also have queuing, so queued message, and you can have a hot key 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

Hot takes

The Claude Code and Codex mobile apps are nowhere near good for working on code remotely, and it's "hilarious" that they pretend otherwise.

His reasoning: He says anyone who has tried to use them for remote work knows they just do not work for it.

If you've tried to fight the Claude Code or Codex mobile apps to work on your stuff remote, you know how bad they are. They just do not work for this. It It is hilarious that they pretend they do. They are so far from good. It's insane.

Other AI coding tools try too hard to put everything in front of you. A good tool should hide work in progress until it needs you.

His reasoning: When running lots of work in parallel, he doesn't want anything in his face unless it needs his attention. Constant visibility overwhelms him, and he cites his ADHD.

don't want things in my 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

The big labs are drifting away from developers. ChatGPT pushes plugins, and Claude Code is trying to be for general work instead of code, which he calls obnoxious.

rest of the world is going the other way, where ChatGPT is trying to shove dots and spaces and plugins in your face, Quad Code's trying to be for work, not for code, which is obnoxious. Everybody else is forcing code into a tab in their other apps. T3 Code is built for code.

How Theo thought about this

Theo is blunt that for months T3 Code was simply worse than the corporate apps. He was half ashamed to promote it and used the Codex app as much himself. It turned around when the team built things the labs didn't care about: the sidebar, remote control and how subscriptions work together.

project. I'm going to be incredibly real and say I was almost a bit ashamed to push it, because I knew what we built wasn't as good as the alternatives. I was using the Codex desktop app as much, if not more, than I was using T3 Code.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 of things. And around July of this year,

Theo expected his opinionated inbox sidebar to be unpopular, so he put it behind hidden flags rather than forcing people into it. It turned out people liked it. He admits switching has growing pains, but says you'll see why once you think about threads his way.

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 hidden flags. I didn't want people to get forced into this cuz I didn't thinkthey 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

Sign in to mark what you've watched and keep notes.

Have a question about this one? Ask about this video →

Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.