Prompt & Done · AIRFIELD

Run three Claude Code agents at once from one free app

T3 Code (pingdotgg/t3code, MIT) is a free, open-source app that runs your coding agents side by side: Claude Code, Codex, Cursor, OpenCode and more, on the logins you already have. Desktop app, local web app, and an iOS / Android app that works as a remote for your computer.

Checked 7 Nov 2026 against the repo README and docs, and our own run on a Mac (T3 v0.1.7, Claude Code 2.1.319, Claude Sonnet 5.5).

1 Install (2 minutes)

First install and log in to at least one agent on the same computer, e.g. Claude Code: claude auth login (or Codex: codex login).

Try it once without installing (needs Node.js):

npx t3@latest

Or install the CLI: curl -fsSL https://t3.codes/install.sh | sh, then run t3. Desktop app: brew install --cask t3-code (macOS), winget install T3Tools.T3Code (Windows), or the .deb / AppImage from GitHub Releases. No t3 binary for Intel Macs: use the desktop app there.

github.com/pingdotgg/t3code · t3.codes

2 Start three agents in parallel

Add your project (Add project → Local folder), pick a Claude model in the model menu, type the first job and press Cmd+Option+Enter: it sends, keeps that thread running in the background and opens a fresh composer. Type job two, same shortcut; job three, Enter. The sidebar's Working group shows every thread live and flips each to Done.

Our test: three jobs in a tiny repo (build a weather page, write + run tests, rewrite the README), all three finished about 27 s after the first was sent; 9/9 tests passed; 5 files changed, +260 −2.

Jobs that touch the same files can collide. For bigger work, pick New worktree under the composer so each agent gets its own copy.

3 Watch it from your phone

Install the T3 Code app (iOS / Android), then on the computer run t3 up and choose T3 Connect (free, up to 3 machines) or local network. The phone is a remote: your computer does the work and has to stay on.

Know before you go: the app is free; you pay your provider as usual. T3's own FAQ says using Claude through the Agent SDK drains the 5-hour limit faster than Claude Code directly (not the weekly limit). It's "very early", so expect bugs.

docs: FAQ · docs: remote access