T3 Code - Minimal Web GUI for Coding Agents

T3 Code is the missing middle between terminal-only coding agents and full IDE-style AI editors.

It does not try to be the model provider. It does not replace Codex, Claude Code, Cursor, or OpenCode. Instead, it gives those coding agents a minimal web and desktop GUI.

That distinction matters. Before T3 Code can do useful work, you still need at least one supported coding agent installed and authenticated.

T3 Code GitHub Source Code T3 Code Website License: MIT

What is T3 Code?

T3 Code is a minimal GUI for coding agents.

The README currently lists support for:

• Codex
• Claude Code
• Cursor
• OpenCode

That means your first setup step is not “enter an API key into T3 Code.” Your first setup step is to install and authenticate a supported agent:

codex login
claude auth login
cursor-agent login
opencode auth login

Then you can start T3 Code:

npx t3@latest

or check the CLI first:

npx t3@latest --help

Repository Snapshot

I cloned the repository into:

tmp/foss-post/t3code

Source state:

Architecture

The architecture docs describe T3 Code as a Node.js WebSocket server that serves a React web app and wraps coding-agent runtimes.

The simplified flow:

Browser React app
  -> WebSocket transport
  -> apps/server
  -> provider service / orchestration engine
  -> coding-agent runtime such as codex app-server

Important repo areas:

The relay is not in the hot path for normal local traffic. The relay docs describe it as a hosted control plane for remote environment discovery, short-lived credentials, mobile notifications, and Live Activities.

Local Field Test

Local environment:

Published CLI Help

Command:

/usr/bin/time -v timeout 90s npx -y t3@latest --help

Result:

One warning appeared:

npm warn EBADENGINE Unsupported engine
package: [email protected]
required: node ^22.22.2 || ^24.15.0 || >=26.0.0
current: node v22.22.0

The warning did not stop the help command.

The CLI help showed:

t3 start
t3 serve
t3 auth
t3 project

Important flags:

--mode web|desktop
--port <port>
--host <host>
--base-dir <path>
--no-browser
--auto-bootstrap-project-from-cwd
--tailscale-serve

Local Server Start

Then I started the server on a high localhost port:

npx -y t3@latest serve --host 127.0.0.1 --port 14173 --no-browser

Observed:

Running all migrations...
Migrations ran successfully
Listening on http://127.0.0.1:14173
T3 Code server is ready.
Connection string: http://127.0.0.1:14173
Pairing URL: http://127.0.0.1:14173/pair#token=...

The server generated a one-time pairing token and QR-style terminal output.

My HTTP fetch loop did not capture the HTML before the process was stopped, so I am treating this as server-start validation, not full browser UI validation.

Useful Keyboard Shortcuts

T3 Code reads custom keybindings from:

~/.t3/keybindings.json

Default shortcuts from the docs:

mod means cmd on macOS and ctrl on non-macOS systems.

Example override:

[
  { "key": "mod+g", "command": "terminal.toggle" },
  { "key": "mod+shift+g", "command": "terminal.new", "when": "terminalFocus" }
]

Docker Notes

I did not find a Dockerfile or Docker Compose deployment path in the main app surface during this pass.

T3 Code is better framed as:

• npx t3@latest for quick local use
• desktop installers for normal app use
• source development through vp / PNPM workspace tooling
• relay deployment only for T3 Cloud internals

Because I did not validate a Docker Compose path, I did not add a Home-Lab compose file.

Search / Positioning Notes

Search results already contain several “T3 Code GUI for coding agents” style posts, plus the official t3.codes site and GitHub repo.

The useful angle for our post is first-hand validation:

• exact repo commit
• exact local command
• exact server output
• dependency warning
• clarification that T3 Code wraps existing coding-agent CLIs
• keyboard shortcuts and provider setup details

T3 Code vs Herdr vs Paperclip

These tools can sound similar because all of them sit around coding agents, but they target different layers:

T3 Code is the local operator interface. It is closer to “make my agent sessions easier to drive” than “schedule autonomous work across a backlog.”

FAQ