The short version

T3 Code is not another coding agent. It is a control surface: a phone app, a web app, and a desktop app that all talk to one server running on your machine, and that server drives the coding agents you already have installed — Claude Code, Codex, Cursor, Grok Build, OpenCode, Google Antigravity, and fx. Seven agents, one front door.

It sits at 52.5k GitHub stars under MIT, ranked #3 on GitHub Trending for the week of Nov 2–8. The author's own warning is refreshingly honest: "We are very very early in this project. Expect bugs."

The problem it actually solves

If you run two or three coding agents at once, your problem is not a missing agent. It is too many front doors: a terminal window per agent, separate sessions, separate permission prompts. Step away from your desk and checking on a long-running agent means SSHing back into your machine first.

T3 Code flips the layering. The agents stay exactly as they are; you add one control plane on top. A server runs on your machine, every client (phone, browser, desktop) connects to it, and from any of them you can watch sessions, send instructions, and approve permissions for all seven agents.

Getting started: three commands 

curl -fsSL https://t3.codes/install.sh | sh t3

t3 starts the server and opens the local web app. To keep it running in the background:

t3 service install

Just want a one-off trial without installing?

npx t3@latest

Desktop packages exist too: winget or scoop on Windows, Homebrew on macOS, .deb for Debian/Ubuntu, AUR for Arch, and nixpkgs for Nix users.

One prerequisite people miss: each agent's own CLI still has to be installed and logged in first — claude auth login for Claude Code, codex login for Codex, agent login for Cursor, and so on. T3 Code is the control plane only. It does not install agents for you, and it does not manage your subscriptions or billing.

Four design choices worth noticing

1. Remote access is native, not bolted on. The repo's docs have a dedicated guide for reaching the server from your phone or another machine. That is the biggest difference from similar tools that only ship a local UI.

2. Multi-account is a documented feature. Codex and Claude each get their own multi-account documentation. If you juggle a personal and a work account, this is considerably cleaner than opening yet another terminal profile.

3. It connects outward over MCP. The docs cover connecting Claude Code, Codex, ChatGPT and other agents over MCP — the control plane itself can be called by other agents. Permission modes get their own doc too, which tells you the authors treat "who is allowed to make an agent do what" as a first-class concern.

4. The pitch is that there is no pitch. The README literally has a section titled "Wait, what are you selling me?" and the answer is "Nothing." The authors built it because they wanted the best agent development experience, judged the existing options (the Codex desktop app, Conductor, Claude Desktop, Cursor Glass) as not meeting their bar, and want you to have everything you need to fork it if they go astray.

The honest limitations

Three things to know before your expectations drift:

  • It is very early, by the author's own admission. Do not bet your main production workflow on it yet.
  • Large feature contributions are (mostly) not being accepted yet — read the contributing notes before you build something big on top.
  • There is no standalone docs site, only the docs/ folder in the repo, and installing/logging into all seven agents remains your homework.

Practical takeaways

  1. If you regularly use two or more coding agents, this is worth an afternoon. Single-agent users will feel less of the benefit.
  2. Get t3 working locally first, then set up phone remote access. Do not start with the always-on service.
  3. For multi-account setups, read the Codex/Claude multi-account docs before writing your own switching scripts.
  4. Treat it as a control plane, not an agent — model choice, subscriptions, and billing still live with each agent itself.

How many coding agents are you running at once — and how are you keeping their front doors straight?

Facts checked against the project's GitHub repo and README on 2026-11-09 (52,526 stars, 10,019 forks, MIT). Star counts move; the architecture point does not.