T3 Code: my Claude and Codex subscriptions in one app, without the nagging worry

I wrote a whole article here about why Pi is my coding harness.

Terminal, few tools, everything pluggable, no foreign opinion in the way. That still holds.

And yet there is now a second app running on my private machine, with a window, a sidebar and buttons.

It is called T3 Code.

The reason is neither comfort nor curiosity. It is a fairly unromantic one: my Claude subscription.

Why I need a second app at all

I said it in the article about my LLMs for Pi: a normal Claude subscription is not a ticket into Pi.

Technically, it works. There are workarounds that slip a third-party harness under the subscription login. But that is not the path Anthropic intended, and if they decide to treat it as circumvention, the worst case is losing the account.

For a few cheaper agent runs, that would be a silly trade.

So until recently I had two options, and I liked neither.

Either pay for Claude through the API, which turns into a bill with personality once you do serious agent work. Or use Claude Code natively in the terminal and switch tools every time I want to switch models.

Privately, I mostly work with Claude and Codex. Two subscriptions, two CLIs, two completely different interfaces. And every time I wanted to swap models in the middle of a task, it meant copying the session, re-explaining the context, starting over.

Not terrible. Just a small annoyance, every single time.

What T3 Code does behind the scenes

T3 Code comes from the team around Theo Browne, is open source under MIT and describes itself as an “agent harness control surface”.

Clunky phrase. Accurate, though.

T3 Code is not an agent itself. It brings no model of its own, no access of its own to Anthropic or OpenAI, no login trick of its own.

Instead it uses what is already installed on my machine: Claude Code and Codex, each as the official CLI. For Claude the setup step is simply claude auth login, for Codex it is codex login. After that, T3 Code drives both and puts an interface in front of them.

That is the part that matters to me.

My Claude subscription still runs inside Claude Code. My Codex subscription still runs inside Codex. T3 Code just sits in front and talks to the tools Anthropic and OpenAI ship themselves. No spoofed client, no rerouted token.

Whether Anthropic sees it differently in a year, I cannot promise. But it is a completely different level of cleanliness compared to what some forums sell you as “works fine, though”.

Claude and Codex are just the two I happen to use. T3 Code now also talks to Cursor, Grok Build and OpenCode, on the same principle: install the vendor’s CLI, log in there, done. Cursor and Grok still carry an “Early Access” label inside the app, and on my machine all three are switched off because I pay for none of them. If you have one of them, it lands in the same sidebar.

And as a side effect it solves a second problem: if you keep separate work and personal accounts, you create one config directory per account and T3 Code treats them as separate environments. One is enough for me, but I know enough people for whom that alone would be the reason to install it.

Switching models without switching tools

The everyday gain is almost boring.

I open a thread, pick Claude or Codex, pick a model, work. For the next thread I pick the other provider. Same sidebar, same composer, same shortcuts, same diff view.

Two apps with two mindsets become one app with two brains.

×

And only now do I notice how often I used to stay with the wrong model out of laziness, because switching was effort. On substance I am the bigger Codex fan, so that is my default. But as soon as frontend work comes up, I fall back to the Claude models every single time. And that fallback is now one click instead of a tool switch.

On top of that come things the providers’ own apps do not solve as well for me:

• Several threads in parallel, visibly. The sidebar shows me which agent is working, which one is waiting and which one is done. In the terminal I used to solve that with tmux and goodwill.
• Permissions are a dial. Four modes from “ask me about everything” to “just do it”, labelled the same for both providers. T3 Code translates them into whatever Claude Code and Codex call them internally.
• Pull requests from inside the app. Title and description get suggested, open reviews land as tabs. Finishing a task no longer means switching to the browser.
• Carrying on from my phone. With T3 Connect I pair my phone to the machine the agents run on. When the agent has yet another question, I no longer have to walk to the desk for it.
• It costs nothing. No licence, no subscription on top. The team writes that they built it because they wanted it themselves.

×

None of this is magic. But together it is enough everyday life that I hardly open the native apps from Anthropic and OpenAI anymore. The models stay the same. The frame around them is, in my opinion, simply more pleasant.

The honest limits

By its makers’ own words, T3 Code is “very very early”. You are told to expect bugs, and you do.

The team is also barely accepting outside contributions right now, because too much is still moving. For me that means keeping in mind that the app is younger than the CLIs underneath it, and not being surprised when something looks different after an update. I still would not hand it to a colleague without that sentence attached.

The second limit is the flip side of the advantage: T3 Code needs the CLIs. Without at least one of them installed, the app is a pretty, empty window. If you do not use those tools already, this is not an entry point, just one more install per provider.

Then there is the fact that it is not Pi. It is nowhere near as open. I do not write an extension that gives the model a new tool or rebuilds the system prompt. T3 Code is a front end for other people’s harnesses, not a harness of its own. If you want to shape your agent, Pi is still the better place.

Which leaves the point I refuse to play down: what a subscription allows is not decided by T3 Code, and not by me. Anthropic and OpenAI decide, and both can change their rules tomorrow. As I read it today, the official CLI behind a third-party interface is the clean path. If you read it differently, please read the rules, not me.

Would I recommend T3 Code?

Yes, to a clearly defined audience.

If you have a Claude subscription or a Codex subscription, work with coding agents and do not want to lie awake thinking about account bans, T3 Code is currently the front end I would show you. Especially if you have both subscriptions and are tired of jumping between them.

If you want to assemble your own harness, stay with Pi or something similar. T3 Code does not want to hand anything over to you. It wants to make the things Claude Code and Codex can already do nicer to operate.

And if you need software that is finished: wait another six months.

For me, it is the first app with a window that I keep open next to the terminal voluntarily.

Pi builds my workflow. T3 Code makes sure I can use the Claude subscription I am paying for anyway with a clear conscience.

Together, that is exactly the mix I wanted.