Free & open source · MIT · 64k+ stars →

One window for
every coding agent.

T3 Code is the open-source control plane for Claude Code, Codex, Cursor and friends. Run them in parallel, review the diffs, ship the PR — from your desk or your phone. Here’s everything you need to learn it.

$npx t3@latest
Claude CodeCodexCursorOpenCodeGrok BuildAntigravityKiroPiMuse CodefxAny ACP agentClaude CodeCodexCursorOpenCodeGrok BuildAntigravityKiroPiMuse CodefxAny ACP agent
What is T3 Code

Not another agent. The cockpit for the ones you already use.

Terminals are great for one agent. They get messy at five. T3 Code runs a small server next to your repositories, starts the official agent harnesses for you, and streams everything to clean desktop, web and mobile clients.

How T3 Code is put togetherDesktop, web and phone clients connect over one authenticated WebSocket to a T3 Code server on your machine. The server runs provider adapters for each agent, plus Git, terminals and a SQLite event store.CLIENTSDesktopmacOS · Windows · LinuxWebany browserPhoneiOS · AndroidWebSocket · typed RPCLAN · Tailscale · T3 ConnectYOUR MACHINET3 Code serverthreads · approvals · checkpointsGit + worktreesTerminalsSQLite logYour reposProvider adaptersYOUR AGENTSClaude CodeCodexCursor · OpenCodeGrok · Antigravityany ACP agent…your subscriptions, your logins
Local first.

State lives in SQLite on the server machine. Clients are thin views — they never run agents or Git themselves.

Official harnesses.

No home-grown agent loop. T3 Code talks to each provider’s own SDK or protocol, so behaviour matches the CLI.

One event model.

Every agent’s output is normalised into one stream, which is why switching or mixing agents just works.

Features

Everything a terminal tab can’t give you.

The pieces that make running several agents at once feel calm instead of chaotic.

Bring your own subscription

T3 Code never resells tokens. It drives the agents you already pay for and are signed in to — no extra keys, no quota caps on top.

Many agents, one thread

Switch models mid-conversation or hand a task from Claude Code to Codex without losing context or leaving the window.

Desktop, web and phone

One local server keeps every client in sync. Start a thread at your desk, approve the next step from your phone.

Worktrees built in

Give each thread its own Git worktree and branch, so parallel agents never trip over each other’s files.

Diff, commit, push, PR

Review diffs per turn, per branch or uncommitted — then commit, push and open a pull request with generated titles.

Built to be fast

Performance is measured and dropped frames are treated as bugs. A GUI that keeps up with agents running in parallel.

Supported agents

Bring your own sub.

Install an agent, sign in once on the server machine, and T3 Code picks it up. Your phone and browser never need the agent installed.

Claude Code
Codex
Cursor
OpenCode
Grok Build
Antigravity
Kiro
Pi
Muse Code
fx
Any ACP agent
Install

Up and running in a minute.

Pick your platform. Then install and sign in to at least one agent, like Claude Code or Codex.

Homebrew
$brew install --cask t3-code
CLI (installs the t3 command)
$curl -fsSL https://t3.codes/install.sh | sh
1Install T3 Code
2Sign in to an agent — e.g. claude auth login
3Open a project and start a thread

Debian/Ubuntu: grab the .deb from the GitHub releases. Phone apps are on the App Store and Google Play.

Permission modes

Decide how much rope each agent gets.

Set it per thread. If a provider doesn’t support a mode, T3 Code falls back to asking.

Supervised

File edits
Ask
Commands
Ask

First runs, unfamiliar repos, learning how an agent behaves.

Auto-accept edits

File edits
Allow
Commands
Ask

Day-to-day coding where you still want to vet shell commands.

Auto

File edits
Allow
Commands
Risky ones ask

Trusted projects; routine actions flow, risky ones pause.

Full access

File edits
Allow
Commands
Allow

Sandboxed boxes and worktrees you can throw away. The default for new threads.

Plan mode is separate: it changes how the agent works (analyse first, then build), not what it’s allowed to touch.

Core concepts

Six words that explain the whole app.

Learn these and every screen in T3 Code makes sense.

01
Environment
One running T3 Code server and the machine behind it. It owns files, projects, threads, credentials and terminals.
02
Project
A directory registered with an environment. Can carry a checked-in t3.json for shared settings.
03
Workspace
Where the agent works: the current checkout, or a fresh Git worktree with its own branch.
04
Thread
A durable conversation for one task, stored on the server and visible from every client.
05
Turn
One cycle from your message to the agent finishing, failing or waiting for you.
06
Checkpoint
Hidden Git refs around each turn that let you revert the chat, the files, or both.
Field notes

Habits that keep parallel agents sane.

  1. Start in Supervised mode with a read-only task to confirm the agent starts, streams and asks for approval.
  2. Use worktrees for truly independent tasks — and keep ports and databases separate between them.
  3. Keep one writer per checkout. Use extra threads for research or review.
  4. Checkpoints are a recovery aid, not a backup. Commit and push work you have reviewed.
  5. Keep the server and every client on the same version; update them together.
  6. Treat pairing links like passwords and never expose a server directly to the public internet.
Guides

Learning paths, coming soon.

Step-by-step walkthroughs we’re writing next. Each one tested on a real machine.

Setup

Your first thread in 10 minutes

Install, sign in to an agent, open a project and ship a small change.

Coming soon
Agents

Claude Code & Codex setup

Sign-in, model picking and handing work between agents.

Coming soon
Remote

Phone approvals over Tailscale

Pair your phone securely and approve steps from anywhere.

Coming soon
Git

Parallel agents with worktrees

Run several tasks at once without merge chaos.

Coming soon
Deep dive

How T3 Code works inside

Event sourcing, Effect RPC over WebSocket and provider adapters.

Coming soon
Compare

GUI vs terminal workflows

When a control plane beats juggling CLI tabs — and when it doesn’t.

Coming soon
FAQ

Questions people actually ask.

What is T3 Code?

T3 Code is a free, open-source (MIT) app for running coding agents such as Claude Code, Codex, Cursor and OpenCode from one interface. It runs a server on your machine and gives you desktop, web and mobile clients on top of it.

Is T3 Code an AI model or a coding agent?

Neither. It is a control plane. The actual work is done by the agents you already have installed and signed in to; T3 Code manages their sessions, threads, Git workspaces and approvals.

How much does T3 Code cost?

Nothing. T3 Code is MIT licensed and has no paid tier. You only pay for whatever agent subscriptions you already use.

Which coding agents does it support?

Claude Code, Codex, Cursor, OpenCode, Grok Build, Antigravity, Kiro, Pi, Muse Code, fx and any agent that speaks the Agent Client Protocol (ACP).

How do I install T3 Code?

On macOS run “brew install --cask t3-code”, on Windows “winget install T3Tools.T3Code”, on Arch “yay -S t3code-bin”. To try it without installing anything, run “npx t3@latest”.

Can I use T3 Code from my phone?

Yes. There are iOS and Android apps. Pair your phone with a one-time link or QR code over a trusted network such as your LAN, Tailscale or T3 Connect.

Does my code leave my machine?

Project state stays on the machine running the server. Prompts, selected context and tool results still go to whichever agent provider you choose, exactly as they would without T3 Code.

What happens if my computer sleeps?

Work stops, because the server and agents run on that machine. For unattended runs, use an always-on machine or a remote server.

Your agents deserve a proper home.

Install T3 Code, connect the agents you already pay for, and run them side by side.