T3 Code Turns Your Phone Into a Remote for Every Coding Agent

Theo Browne's new repo answers a question with a question. Under a heading that reads "Wait, what are you selling me?" the T3 Code README says: nothing.
That is either refreshing or a warning sign, depending on how many early-stage developer tools you have installed. T3 Code launched into roughly 39,700 stars with about 6,900 added in the past week, and its one-line description is "agent harness control surface" — which is accurate and almost meaningless until you see it running.
The actual idea
You already have coding agents on your machine. Claude Code, Codex, Cursor CLI, Grok Build, OpenCode, Google Antigravity, fx. Each has its own terminal, its own permission model, its own way of telling you it is waiting for approval.
T3 Code does not replace any of them. It sits on top and gives you one client for all of them: an Electron desktop app, a web app at app.t3.codes, and iOS and Android apps. The server runs on your computer. Your phone is a remote control.
If you have ever walked away from a laptop while an agent was half-way through a refactor, and come back to find it blocked on a yes/no prompt, the appeal is immediate. The agent keeps working; you approve from the couch.

Install, and the several ways to do it
The command line route is a one-liner:
curl -fsSL https://t3.codes/install.sh | sh
On Windows in PowerShell:
irm https://t3.codes/install.ps1 | iex
Then t3 starts the server and opens the local web app. t3 service install keeps it running in the background, t3 update moves you to a newer release, and npx t3@latest runs it once without installing anything.
The desktop app has more packaging than most projects this young bother with: winget install T3Tools.T3Code, scoop install extras/t3code, brew install --cask t3-code, a .deb for Debian and Ubuntu, yay -S t3code-bin on Arch, and nix profile add github:NixOS/nixpkgs/nixos-unstable#t3code.
Before any of it works you need at least one provider authenticated on that machine. Codex via codex login, Claude via claude auth login, Cursor via agent login, and so on. T3 Code rides your existing subscriptions. There is no T3 subscription, which is the "what are you selling me" answer.
Why "control surface" is the right phrase
Because the interesting design decisions are all about surfaces, not models.
There is a documented permission-mode system, keyboard shortcuts, per-project settings, and a remote access guide. There is a way to connect outside agents — Claude Code, Codex, ChatGPT — over MCP. There is source control integration. There is multi-account support for Codex and Claude, which matters more than it sounds if you have a work account and a personal one and keep switching shells.
And there is a background service mode, which is what turns this from a toy into something you leave running. The architecture is: server on the machine that has the repositories and the credentials, clients anywhere.
That architecture is also the thing to be nervous about. You are exposing an interface that can approve tool calls and run commands in your repositories, reachable from a phone. The project documents permission modes and an outside-agents MCP path, but "read the docs before you expose this" is not paranoia, it is the minimum.
What the docs actually cover, and what that tells you
The documentation set is short but specific, and every entry is a decision someone had to make: permission modes, keybindings, project settings, appearance, remote access, connecting outside agents over MCP, keeping app and server in sync, source control integrations, multiple accounts per provider, and running as a background service.
That is not a feature list. It is a list of the things that go wrong when you let a remote client drive a local agent. "Keeping app and server in sync" exists because someone shipped a version mismatch and watched the client send garbage. "Permission modes" exists because the default answer to "should this run without asking" cannot be the same for a scratch repo and a payments service.
My advice, if you try it: set the permission mode before you connect a phone, not after. The tempting order is to get it running on the couch first and think about approvals later, and that order is how people end up with an agent that has been quietly approving its own rm commands for a week.
The honesty is load-bearing
Two paragraphs in the README do more for my confidence than any feature list.
The first: "We are very very early in this project. Expect bugs."
The second: "We are (mostly) not accepting contributions yet. Small fixes may be considered. Big features will not be."
Most projects at 39,000 stars are drowning in pull requests and pretending otherwise. Saying no early keeps the architecture coherent, and the README pairs it with the reason you should still trust the thing: if the direction goes wrong, you have everything you need to fork it and build the editor you want. That is a real commitment for a project that ships an Electron app, a CLI, and two mobile clients under a permissive posture.
There is also a delightful bit of infrastructure trivia buried in the contributing section: T3 Code uses Vite+, so you need the global vp tool, installed via curl -fsSL https://vite.plus | bash, and dependencies via vp i. Whether a new package manager for Vite is a good idea is a separate argument, but it tells you this team ships on its own stack.
What I am not convinced by
The mobile angle is the marketing, but I suspect the desktop app is where the real usage lands. Approving from a phone is great for long-running tasks and terrible for anything that needs you to read a diff. Scrolling a 400-line patch on a 6-inch screen while standing in a kitchen is not a workflow, it is a compromise.
The deeper question is whether a control surface is a durable product or a transitional one. Right now every agent vendor wants its own client, because the client is where permission, context and billing live. A neutral layer on top is valuable precisely while that fragmentation exists. The moment one vendor ships a genuinely good cross-agent client — and Anthropic and OpenAI both have the incentive — the neutral layer's job gets harder.
Which is presumably why the README insists on being forkable. If the value is the surface, the surface has to be yours.
Try it if
You run more than one agent, you leave long tasks running, and you want one place to see what all of them are doing. That is the exact audience, and for that audience the install is five minutes and the payoff is immediate.
Do not try it as your primary editor yet. It is not one, it says so, and you will be annoyed when it behaves like software that is three months old.
The interesting thing about this release is not the app. It is that someone looked at the current state of coding agents and concluded the missing piece is not a better model or a better prompt. It is a switchboard.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.