T3 Code quickstart

Get T3 Code running in about two minutes — install a provider CLI, launch with npx t3, pick a runtime mode, and ship your first agent thread.

Last reviewed 2026-07-27 · verified against the pingdotgg/t3code repo

There are only two hard requirements: a provider CLI that is installed and logged in, and a recent Node.js. Everything else T3 Code handles for you.

Step 1 — Install and authenticate a provider#

This is the step people skip, and it is the cause of most first-run failures. T3 Code does not ship an AI model and cannot log in for you; it launches your provider's CLI as a subprocess.

Pick one:

# OpenAI Codex — the most mature integration
npm i -g @openai/codex
codex login

# Anthropic Claude Code
npm i -g @anthropic-ai/claude-code
claude auth login

# Cursor agent
cursor-agent login

# OpenCode
opencode auth login

Verify it works on its own before involving T3 Code:

codex --version

Step 2 — Launch T3 Code#

The fastest path needs no installation at all:

npx t3@latest

This downloads the current release, starts a local Node server on port 3773, serves the React front end, and opens your browser at http://localhost:3773.

Prefer a real application in your dock? Install the desktop build instead — it bundles its own runtime, so you don't need Node at all:

# Windows
winget install T3Tools.T3Code

# macOS
brew install --cask t3-code

# Arch Linux
yay -S t3code-bin

Full per-platform notes, including .exe / .dmg / .AppImage downloads, are on the installation page.

Step 3 — Choose a runtime mode#

On first launch you pick how much freedom the agent gets:

Start on Supervised for your first session. You will learn far more about what the agent is actually doing, and you can switch later. Runtime modes explained →

Step 4 — Add a project#

Point T3 Code at a git repository on your machine. It reads the working tree directly — nothing is uploaded to a remote service.

Step 5 — Run your first thread#

Create a thread and give it something small and verifiable:

Create a file called hello.js that prints "Hello from T3 Code!" and nothing else.

Watch for three things, because they tell you the whole system is wired correctly:

1. Streaming output — tokens appear live, which means the WebSocket is connected.
2. A file appears in the diff panel — which means the agent has real filesystem access.
3. The git controls light up — which means the repo was detected and worktree mapping works.

If output streams but no diff shows, the agent is talking but cannot write. That is almost always a runtime mode or permissions issue rather than a T3 Code bug.

Step 6 — Review, commit, ship#

Once the agent finishes, the controls in the top right take you through the whole git flow without touching a terminal: inspect the diff, commit, push, and open a pull request. Large diffs collapse by default in recent builds; there is a collapse-all toggle in the diff panel.

For the parallel-agent workflow — a worktree per thread, three branches at once — read worktrees, diffs and pull requests.

When it doesn't work#

In rough order of how often they come up:

Everything else lives in troubleshooting.

FAQ#

How long does T3 Code take to set up?#

About two minutes if a provider CLI is already installed and authenticated. Budget five to ten minutes if you also need to install and log into Codex or Claude Code for the first time.

Do I need to install anything to try T3 Code?#

No. Running npx t3@latest downloads and runs the current release without a permanent install. You do still need a provider CLI such as Codex installed and logged in separately.

What port does T3 Code use?#

Port 3773 by default, serving both the web UI over HTTP and the agent session over WebSocket at ws://localhost:3773.

Should I use Full Access or Supervised mode?#

Start with Supervised so you can see every command before it runs. Switch to Full Access once you trust the setup and the repository, since approval prompts slow down long agent runs considerably.