T3 Code is a remote control surface for coding agents: a phone, web and desktop client that drives the agent command line tools already installed on your own computer. It is aimed at developers who kick off a long task with Claude Code, Codex or Cursor and then want to leave the desk without losing the thread. The project supplies no model and no intelligence of its own, only the interface and the plumbing between a phone and the agent session running on a machine you own.
What it does
T3 Code puts a front end on agent harnesses that normally live in a terminal. The README says it works with your existing subscriptions on six of them, as long as they are installed and authenticated on the machine:
- Claude Code
- Codex
- Cursor
- Grok Build
- OpenCode
- Google Antigravity
The same work is reachable from three places: native apps for iOS and Android, a hosted web app at app.t3.codes, and an Electron desktop app from t3.codes. The video's framing is the simplest description of the point: start a task at your desk, walk away, and keep steering it from your phone.
How it works
You install a small program on the machine where your agents live. Running t3 starts a server there and opens the local web app; the README also documents t3 service install to keep that server running in the background. The clients talk to that server, and the server drives the provider CLIs using the logins already sitting on the machine, so the billing relationship stays between you and each provider.
Antigravity is handled differently. The README says you enable it in Settings and then use the "Install Antigravity" button and "Sign in with Google", with no CLI required.
The repository is TypeScript under the MIT license. The authors say the openness is deliberate: if the project ever goes the wrong direction, they want users to have everything needed to fork it and build the editor they want.
Getting started
You need at least one provider installed and logged in before T3 Code is useful, and the README is explicit about it, listing codex login, claude auth login, agent login for Cursor, grok login and opencode auth login. Beyond that the project does not list hardware requirements: no chip, memory or GPU target appears in the materials, because the heavy work is done by the agent CLIs and the cloud models behind them.
T3 Code itself is free and MIT licensed, and the authors answer the "what are you selling me" question with "nothing". The intelligence is not free. You pay whichever provider you point it at, through the Claude, Codex, Cursor, Grok or Google plan you already hold, so this is an extra interface on top of a subscription rather than a replacement for one.
Install is a single piped command. On macOS and Linux:
curl -fsSL https://t3.codes/install.sh | sh
On Windows, in PowerShell:
irm https://t3.codes/install.ps1 | iex
The clients are not built from a clone. The iOS app is on the App Store and the Android app on Google Play (both linked from the README), the web app is at app.t3.codes, and the desktop build is at t3.codes. Source lives at github.com/pingdotgg/t3code.
When to use it / when not
The main catch is maturity. The repository was created in February 2026 and is still being pushed to daily, and the video covering it says plainly to expect bugs. The second catch is structural: your machine has to be awake with the server running, because there is no cloud agent behind this, and any harness that is not already authenticated locally simply will not show up.
Use it instead of opening an SSH session from a phone terminal when you want a real touch interface for approving steps, reading output and nudging a running agent, rather than fighting a terminal with a mobile keyboard. Don't use it if you only ever work at one machine in one window: a plain terminal, or the vendor's own desktop app, already covers that case and adds no moving parts.
Alternatives
The README names its own points of comparison and calls them inspirations that did not meet the authors' bar: the Codex desktop app, Conductor, Claude Desktop and Cursor Glass. What T3 Code claims over them is breadth plus reach, one surface across six harnesses with mobile treated as a first-class client. That is the authors' own claim, not an independent assessment, and the channel does not run the tool itself.
If you run agents on long tasks and resent being chained to the desk while they churn, this is worth a serious look: the install is two lines, the license is permissive, and nothing about it asks you to change models or move your work to someone else's servers. GitHub's own counters show roughly 31,000 stars and 8,400 forks within months of the first commit, which says how widely the itch is felt, not that the software is finished. Developers who need stability above all, or who hold no agent subscription at all, should let it settle for a few more releases.