t3code: a Chinese translation of the agent harness control panel, and what it actually gets you

T3 Code is a control surface for coding agents you already run on your own machine. The upstream project is pingdotgg/t3code, 36,095 stars, written in TypeScript, MIT licensed. There is now a Chinese translation of its documentation at yangshun2005/t3code-cn, published 2026-11-01.

That translation is the subject here. Not because a translated README is remarkable in itself, but because of what the original project asks of its reader before anything works.

What T3 Code is, precisely

The README calls it an "agent harness control surface." Read that literally. T3 Code ships no model. It does not proxy your API keys to some inference endpoint. It connects to agents already installed and authenticated on your local machine, and gives you an interface to drive them from an iOS app, an Android app, a web app, or an Electron desktop build.

The supported providers, per the README:

• Codex
• Claude Code
• Cursor
• Grok Build
• OpenCode
• Google Antigravity
• fx

If none of those are installed and logged in on your machine, T3 Code has nothing to control. That constraint shapes the entire onboarding flow, and it is the first thing the Chinese version front-loads.

Why the translation matters more than it sounds

The original README is honest about the project's state: "very, very early, expect bugs," and contributions are largely closed — small fixes might be considered, large features will not. So this is not a project where you read the docs, file an issue, and wait. You read the docs, get it running, and decide for yourself.

Getting it running means seven different authentication commands across seven different CLIs, plus platform-specific install paths for macOS, Linux, Windows, Debian, Ubuntu, Arch, and Nix. Every one of those steps is a place where a non-English reader stalls — not because the English is hard, but because the failure modes are terse and the commands are easy to mistype.

The Chinese version addresses this with a structure the original does not have. It separates two things:

1. A 中文快速上手 section at the top, explicitly labeled as supplementary orientation material, not part of the original.
2. The translated README body below it, which aims to stay faithful to the source.

That split is the right call. A translation that silently adds advice is a translation you cannot trust. A translation that only translates leaves the reader to assemble the setup sequence themselves. Marking the boundary lets you use the fast path and still know which parts are the author's additions.

The shortest path to a running instance

The quick-start section condenses the setup into five steps. Here is what it says, in order.

First, install and log in to at least one supported agent. The commands, as given:

# Codex
codex login

# Claude Code
claude auth login

# Cursor CLI
agent login

# Grok Build
grok login

# OpenCode
opencode auth login

Antigravity is the exception — no CLI. You enable it in Settings, then click Install Antigravity and Sign in with Google.

fx has its own wrinkle: fx login must run on the machine hosting the T3 Code server, not just any machine. The Chinese version calls this out explicitly in its pitfalls section. The original mentions it too, but buried in a provider list; the translated version surfaces it as a standalone warning.

Second, install the CLI. macOS and Linux:

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

Windows PowerShell:

irm https://t3.codes/install.ps1 | iex

Third, run t3. That starts the server and opens the local web app.

Fourth, optional service management:

t3 service install   # run in background
t3 update            # upgrade
t3 --help            # full command reference

Fifth, if you want to try without installing: npx t3@latest.

Desktop builds are also available from GitHub Releases or via package managers — winget install T3Tools.T3Code and brew install --cask t3-code are two the README lists.

The network caveat the original doesn't discuss

The Chinese quick-start adds a note that the install scripts and desktop packages all download from overseas addresses: t3.codes, GitHub Releases, and various package managers. Under restricted network conditions you may need your own acceleration or mirror.

It also states plainly that the original provides no mirror sources. That is accurate and worth saying. A translated doc that pretends the network is uniform is less useful than one that flags the gap and admits it has no fix.

What the translation does not claim

No hardware requirements. No minimum OS versions. The Chinese version says so directly rather than inventing them: "原文未说明" appears in the environment requirements section for both hardware and OS version floors.

That restraint is the thing I'd point to if you're evaluating whether to trust this translation. It would have been easy to fill those blanks with plausible guesses. The author didn't.

The one build-from-source detail it does add: you need the global vp (Vite+) tool installed, then run vp i. That comes from the original's build instructions, just relocated to where a reader would look for it.

Who this is for

If you already use Claude Code or Codex locally and want a unified, remote-capable interface across phone, web, and desktop, T3 Code is aimed at you. If you have none of the seven supported agents installed, there is no partial mode — the project simply won't do anything.

The Chinese version lowers the entry barrier for that first group by removing the assembly work: the provider matrix, the per-platform install commands, and the ordering constraints are all in one place, in the language you're reading. My sense is that most of the friction in early-stage projects like this is not conceptual difficulty but sequence confusion — which command first, which machine, which login. That's exactly what the quick-start section targets. Your mileage may vary depending on which provider you pick; the fx path in particular has that server-machine constraint that's easy to miss even when it's written down.

The translation is at github.com/yangshun2005/t3code-cn. It carries zero stars as of publication — it went up today. The upstream project it documents has 36,095. Corrections to the translation go through Issues or PRs, per the README.

Chinese documentation: https://github.com/yangshun2005/t3code-cn

I'm an independent enterprise AI delivery consultant — private LLM deployment, RAG, agent workflows, AIOps. If your team is stuck somewhere in that stack, my inbox is open.