It doesn't sell models; it just takes over the coding agents already logged in on your computer, gaining 3,100 stars in a day

No. 4 on the GitHub AI daily chart, up 3,100 stars in a day, with 38,000 total stars. The project is called pingdotgg/t3code. The author, pingdotgg, is Theo of t3.gg, and create-t3-app also came from him.

In one sentence, here's what it does: it provides a control center for the coding agents on your local machine.

There are plenty of coding agents now. Claude Code, Codex, Cursor, Grok Build, OpenCode, and Antigravity each have their own client, their own subscription quota, and their own place in the terminal or editor. Once you use two or more at the same time, switching costs appear. You don't know which one is still running, which one is stuck waiting for your confirmation, and which one has finished its changes and is waiting for you to review.

t3code's approach is interesting: it doesn't ship its own models. It runs a t3 service locally, connects to the tools already logged in on your computer, and uses your existing subscription quota. Connect from desktop, web, or the mobile app, and you can see what these agents are doing and give them new instructions.

Someone in the comments said that not selling models and only connecting to local logins is a fierce move. Why is it so sharp? This layer doesn't touch compute or tokens; it only does orchestration and observation, so the cost structure immediately becomes different.

There are roughly three things it can do:

• Manage multiple coding agents at once in one interface, without jumping back and forth between terminals
• Check progress remotely from the mobile app when you're out, and send new instructions along the way
• Keep it running as a background service so agents on your computer are always ready

The barrier to entry isn't high. First install and log in to at least one supported agent tool, then use a one-line install script, or try npx t3@latest. The author puts it plainly too: the project is still very early and will have bugs. The license is MIT.

There is a point here that's easy to overlook. The value of this kind of control center has little to do with how strong the models are. It can exist precisely because there are several model providers. No one wants to be locked into a single model, but each has its own strengths and quotas, and that's where the space for a middle layer lies. Model vendors are unlikely to do this kind of work, because its premise is admitting that you won't use only one vendor.

Don't expect an early-stage project to be stable. Before really making it your daily driver, run it on a small project for a couple of days and see how well it restores sessions with the backend agents.

Someone in the comments also asked a more practical question: can sessions from different agents be grouped together by project? When you come back a few days later to find a feature that's half-finished, a normal person's sequence is to think of the project name first, then which tool they were using at the time. Today these control centers are basically organized by tool; organizing by project is another matter. Whoever builds it first will have one fewer complaint to deal with.