T3 Code
An agent harness control surface: keep the agents and subscriptions you already have — Claude Code, Codex, Cursor, Grok Build, OpenCode — and drive them from a fast desktop app, a web app, or a genuinely good mobile app.
Operator's take
The agents live on your machine. You, increasingly, don't. T3 Code's bet is that the missing piece isn't a better agent or another harness — it's a control surface: one server on your computer, and fast apps (desktop, web, and above all mobile) for driving the agents that are already installed there, on the subscriptions you already pay for. "If they're set up on your computer, T3 Code can control them" is the whole pitch in one line, aimed at Claude Code, Codex, Cursor, Grok Build, and OpenCode.
Getting started is one command — npx t3@latest runs the backend and the local web app; desktop apps ship through GitHub releases, winget, Homebrew, and the AUR. The team behind it — T3 Tools, from Theo Browne's create-t3-app orbit; Browne is a well-known dev YouTuber, and the project launched to that audience, which is a real adoption advantage in a lane full of one-repo tools — built it because the existing answers (the Codex desktop app, Conductor, Claude Desktop, Cursor Glass) each missed their bar for performance, remote access, or openness, and they publish it under MIT with an explicit fork-it-if-we-stray stance. The performance claim is the one users repeat most: an Electron app that doesn't feel like one, with the phone experience as the hero story — real testimonials describe doing actual development from a phone two thousand miles from the machine.
The catch is that the project tells you itself, in the README's own words: "We are very very early in this project. Expect bugs," and contributions are mostly not being accepted yet. Documentation lives in the repo rather than a docs site. It's a control surface, not an orchestration layer — it steers the agents you have and leaves worktree choreography and review depth to heavier tools — and it presumes the agents are already set up, because controlling them is the job. Adopting T3 Code today means adopting a fast-moving alpha from a team that ships in public.
What it's good at
- Drives your existing subscriptions — Claude Code, Codex, Cursor, Grok Build, OpenCode; no new agent, no new bill, just remote control of what's installed.
- Genuinely mobile — iOS and Android apps whose users describe doing real development from a phone, including iOS work, from another country.
- Three surfaces, one server — Electron desktop app, web app at app.t3.codes, and mobile, all against the one backend you launch with
npx t3@latest(Node 22.16+). - Fast for its class — performance is the most-repeated compliment from users ("proof electron apps don't have to suck"; "the first coding gui that didn't suck performance wise").
- Open and forkable — MIT-licensed, with an explicit commitment that if the project goes the wrong direction you have everything needed to fork it.
- Linux treated as first-class — AUR packages (stable and nightly), a background-service mode, plus winget and Homebrew for Windows and macOS.
What it's not
- Not finished — the README's own words: "very very early… Expect bugs"; small fixes may be accepted, big features won't be yet.
- Not a harness itself — it controls agents you've already set up; it doesn't ship one of its own.
- Thin on docs — no docs site yet; documentation lives as markdown in the repo.
- Not a review environment — a steering surface for sessions rather than a worktree-and-diff workflow; deeper review flow belongs to heavier tools.