T3 Code 编程智能体控制台

它是什么

T3 Code is an app for controlling the coding agents that are already on your computer. The README says it works from a desktop app, a web app and a mobile app for iOS and Android, using your own subscriptions. It has no model or agent of its own, and the FAQ says it does not work without one. You install Claude Code, Cursor or another supported agent and sign in there, or connect Codex with your ChatGPT plan and Antigravity with Google from T3 Code’s settings. T3 Code then gives you threads on your projects, a permission mode for each thread, the changes each turn made and pull requests, all in one place.

1. Sign in to an agent on the machine that does the work
2. Start T3 Code there, as the desktop app or the t3 command
3. Enable the agent in Settings and start a thread
4. The agent works on that machine, asking as the mode says
5. A phone or browser connects to watch and steer the thread

适合谁

Developers who already pay for Claude Code, Codex, Cursor or another supported agent and want one app for all of them. It fits best if you run agents on a spare computer or a VPS and want to steer them from your phone. Nothing is for sale: the README answers “what are you selling me?” with “Nothing.”

Skip it if you want to run local models, which the FAQ says T3 Code does not support. Skip it too if you have no plan or account with a supported agent.

用在哪里

T3 Code runs on a desktop, in a browser and on a phone, and it drives these agents.

• Desktop app: Windows, macOS and Linux, from GitHub Releases, winget, Scoop, Homebrew, a .deb, the Arch User Repository or Nix
• The t3 command: macOS, Linux and Windows, with no build for Intel Macs
• Web: the local web app that t3 opens, or the hosted one at app.t3.codes
• Phone: the iOS and Android apps, which connect to your computer
• Agents: Claude Code, Codex, Cursor, Grok Build, OpenCode, Google Antigravity and fx in the README, plus Kiro, Pi and Muse Code on the Install page

如何设置

Get it saves a zip of Minda’s setup notes in your downloads folder. It holds a file each for the command line, the desktop app, phones and remote access, Claude Code, Codex, Cursor and the other agents, taken from T3 Code’s docs. Install and sign in to at least one agent on the machine that will do the work, before your first thread. Codex through ChatGPT and Antigravity are set up from T3 Code’s Settings → Providers instead.

1. Install an agent such as Claude Code or Codex and sign in to it, for example with claude auth login or codex login.
2. Install T3 Code with one of the lines below, or with the desktop app.
3. Run t3. It starts the server and opens the local web app.
4. Open Settings → Providers, select the environment and enable the agent. If it does not show up, check that its CLI is on the PATH, or set its Binary path.
5. To reach the machine from another device, run t3 up on it.

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

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

The first line is for macOS and Linux and the second for Windows PowerShell. Both run a script from t3.codes, so read it first if you want to see what it does. To try T3 Code once without installing, run npx t3@latest. The desktop app has its own installs.

The Debian line needs the .deb from GitHub Releases in the current folder. The README also gives scoop install extras/t3code for Windows and a nightly build for Arch.

如何使用

Open a project, start a thread, pick the agent and its model, and send your task. A permission mode decides how often the agent stops to ask. A new thread names its mode under the composer. Press mod+shift+a to change it and mod+shift+m to change the model, where mod is Command on macOS and Ctrl on Windows and Linux.

The modes differ a little by agent. Muse Code and Kiro offer only Supervised and Full access, and fx has only Supervised and Auto. Grok has no Auto-accept edits and Pi has no Auto. On OpenCode and Antigravity, Auto falls back to asking.

示例

You have a Linux box at home with Claude Code signed in, and you want to steer it from your phone. On the box, you install T3 Code and run t3 up.

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

t3 up asks how your other devices reach the machine. You choose T3 Connect and follow its sign-in instructions. On Linux it then keeps T3 Code running as a background service. On the phone, you install the app, sign in to the same T3 Connect account and choose the machine. You open a thread on a project, pick Claude and type the task. The agent works on the box and the thread shows on your phone. The box has to stay on.

新手与专家

新手：Claude Code and Codex are coding agents that you sign in to with their own commands, such as claude auth login and codex login. T3 Code is the app around them and holds no model. The FAQ gives a picture. The model is the chef, the agent program is the kitchen, and T3 Code is the restaurant that lets you bring whatever kitchen you want. So set up at least one agent before your first thread. The Install page says you can launch T3 Code first and configure providers afterwards. A thread is one conversation with an agent, usually on one project.

专家：The FAQ says T3 Code drives Claude through Anthropic’s Agent SDK. A second login is a second provider instance. Claude takes its own CLAUDE_CONFIG_DIR, and each account’s conversations stay apart. Codex takes a shadow home under the same CODEX_HOME, so work and personal accounts keep their own login and continue the same threads. An API key or OpenRouter works too, as another instance with its own keys and endpoint.

Other agents can drive T3 Code in turn. Each environment has an MCP server at its address followed by /mcp. An agent that supports remote MCP with OAuth signs in once, and you choose what it may do. Read only is the default.

claude mcp add --transport http t3 https://<environment-address>/mcp
claude mcp login t3

On the host, t3 thread works without a sign-in. A send with --wait prints the reply. It exits 1 when the run fails or stops, and 2 when the agent waits for an answer you give in the app.

id=$(t3 thread start . "Find why the tests are slow")
t3 thread send "$id" "Now fix it" --wait

A scheduled task can start on a webhook, such as a new pull request or a failed CI job. A public webhook URL needs a T3 Connect managed tunnel. The docs say to turn on Require signature for GitHub, because without one anyone with the URL can start a run with any content.

注意事项

Most of T3 Code runs on your machine, and a few paths reach out.

• Early software: the README says the project is very early, to expect bugs, and that it mostly does not accept contributions yet
• Cost: it is free and MIT licensed, and you still pay your agent’s provider. T3 Connect is free for now for up to 3 machines per account. A paid T3 Chat plan raises that to 10 with Pro or 100 with Premier
• Permissions: the initial default for new threads is Full access, which allows commands and edits without prompts. Change it under Settings → Threads → New threads → Permissions
• Claude limits: the FAQ says Claude usage through the Agent SDK drains the 5-hour limit faster than Claude Code on its own, and that it has been reported to Anthropic
• Where code goes: the FAQ says your code and prompts go from your machine to your provider, not to T3. T3 Code does send product usage events with hashed identifiers, and the phone apps send crash reports. Setting T3CODE_TELEMETRY_ENABLED=false in the server’s environment before it starts turns off the events and relay tracing. Crash reports have their own switch in the phone app’s Settings > About T3 Code > Diagnostics
• Local network: after t3 up --lan, a pairing code is all that keeps other devices on your network out. Use a fresh one-time link for each device
• Versions: an app newer than its server is blocked until you update the machine named in the notice
• Platforms: there is no t3 build for Intel Macs, and Windows has no background service
• License: MIT. The LICENSE file reads “Copyright (c) 2026 T3 Tools Inc.”