Using Codex with T3 Code

Set up OpenAI Codex as your T3 Code provider — install, login, how codex app-server fits in, and why keeping the CLI current prevents most session failures.

Last reviewed · verified against the pingdotgg/t3code repo

Codex was the original target and remains the best-tested integration. If you're evaluating T3 Code and have a choice, start here.

Setup#

bash
npm i -g @openai/codex
codex login
codex --version

That third command isn't optional ceremony. T3 Code launches this exact binary — if it doesn't respond here, nothing downstream will work.

Then:

bash
npx t3@latest

How the integration works#

Per session, T3 Code runs codex app-server and speaks JSON-RPC over stdio to it. Structured events stream back and become WebSocket messages for the UI.

The implication that matters: the app-server protocol is a compatibility surface between two independently-versioned tools. T3 Code updates nightly; Codex updates on OpenAI's schedule. When they drift, the handshake fails.

Architecture in full →

The three Codex problems worth knowing#

Session hangs on startup#

Timed out waiting for initialize — outdated Codex CLI. Update it. Full fix →

'codex' not found when it demonstrably exists#

codex --version works in your shell, T3 Code disagrees. PATH inheritance, not a missing binary. Most common on macOS with nvm or Homebrew, and on Windows after a global npm install without re-login. Full fix →

failed to start codex app-server#

Reported on Windows, sometimes as os error 3. Path resolution or spawn failure, often a stale npm global prefix or a broken shim. Also check 127.0.0.1:9234 — Codex uses it for remote connections and a conflict there blocks bootstrap. Full fix →

Codex CLI vs Codex app vs T3 Code#

Three things share the name, which causes real confusion:

What it is
Codex CLIThe terminal agent. T3 Code requires it
Codex appOpenAI's own GUI for Codex
T3 CodeA third-party GUI that drives the Codex CLI

T3 Code and the Codex app are alternatives to each other; both sit on the CLI. The comparison is on T3 Code vs the Codex app.

Cost#

You pay OpenAI, through whatever your Codex CLI is authenticated with — a subscription or API usage on your key. T3 Code adds nothing. Note that agentic runs consume a lot of tokens: an agent exploring a codebase reads far more than a chat conversation does, so watch usage during your first week.

Getting the most out of it#

  • Use worktrees so several Codex threads run in parallel without colliding
  • Start on Supervised mode to learn what Codex actually does between messages
  • Roll back to a checkpoint rather than arguing with a thread that's gone wrong — it's faster
  • Update the CLI weekly, especially if you're on T3 Code nightlies

FAQ#

Do I need the Codex CLI to use T3 Code?#

You need at least one supported provider CLI. Codex is the most mature option and requires npm i -g @openai/codex followed by codex login before T3 Code can start a session.

What is codex app-server?#

It's the server mode of the Codex CLI. T3 Code launches it per session and communicates over JSON-RPC on stdio, translating the structured events it returns into WebSocket messages for the UI.

Why does T3 Code time out waiting for initialize?#

Almost always an outdated Codex CLI whose app-server protocol no longer matches what T3 Code expects. Updating Codex to the latest version resolves it.

Is T3 Code better than the official Codex app?#

They're different trade-offs. T3 Code adds git worktree isolation, a diff viewer, one-click pull requests and support for multiple agents. The Codex app is first-party and tracks Codex features immediately.