{"slug": "environment-update-a-t3-code-tokenmaxing-constellation", "title": "Environment Update: A T3 Code Tokenmaxing Constellation", "summary": "A developer installed the latest T3 Code nightly build and its TestFlight mobile app from pingdotgg, then ran agents across an M1 Mac mini, an M3 Max laptop, a lightweight terminal laptop, and an iPhone, using ChatGPT and Claude subscriptions rather than API keys. The setup uses the nightly channel everywhere, installs Codex CLI, Claude Code, and GitHub CLI 2.81+, and runs `t3 connect --headless` with `t3 service install` on each worker machine. The author said the configuration maxed out all available hardware and token subscriptions and praised T3 Code for matching Codex Desktop features while working across all subscriptions with strong mobile support.", "body_md": "## 🚀 Maxed out\n\nEarlier this year I [wrote](https://www.bounds.dev/posts/teaching-claude-code-to-drive-cmux/) about how I customized my development environment to automate workflows using agents and [cmux](https://cmux.com/). Since then I transitioned to the [Codex Desktop App](https://chatgpt.com/codex/) for most of my work, and it’s been great.\n\nBut, I’ve had this nagging feeling that by using one app I’m locking myself into a single vendor and not learning as much as I could.\n\nMeanwhile, the rest of the world is *tokenmaxing*.\n\n**This weekend it was my turn!** I installed the latest [T3 Code nightly](https://x.com/theo/status/2106581208193003977) and companion [TestFlight mobile app](https://x.com/theo/status/2106590675106230321), and a few hours later maxed out all of my hardware *and* my tokens.\n\nThat includes Claude and Codex subs, a Mac mini, a big laptop, and a smol laptop that is mostly just a terminal client so I can manage work on the other machines wherever I am.\n\n**I could not be more impressed with T3 Code.** It…\n\n- ✅ has everything I loved about Codex Desktop,\n- ✅ works across all subscriptions, and\n- ✅ has incredible mobile support.\n\nIf you have not experimented with a setup like this yet, **you are missing out**. So, I’m sharing some notes here about how I got things set up, and maybe this roadmap will be helpful for you and your agents, too.\n\n## 🧰 The parts list\n\nAlthough token subs can cost you, the great thing about building your own tokenmaxing constellation is that **almost any *hardware* works**: the Mac mini you bought for OpenCode and gave up on, a cheap old laptop, whatever works.\n\nHere’s what I’ve got:\n\n- 🔌 **M1 Mac mini:** Not powerful, but always on and good enough for lots of tasks, including Xcode mobile app building. This is my default build environment.\n- 💪 **Big laptop:** A pretty powerful M3 Max with lots of memory, good for harder jobs but not always online.\n- 🪶 **Smol laptop:** Light and almost disposable! Mostly just there to run the T3 Code client so I can drive the T3s on other environments where they stay running, and I can disconnect and*close the lid* anytime.\n- 📱 **Phone:** An iPhone running[T3’s V2 mobile client](https://github.com/pingdotgg/t3code/blob/main/docs/user/install.md#mobile-app) .\n- 🌙 **[T3 Code nightly build](https://github.com/pingdotgg/t3code)**\n- 💳 **Subscriptions:** ChatGPT and Claude.\n\n## 🧑🏻🍳 Sharing a recipe\n\nHere are bare minimum instructions to set this up yourself. **Hand them to Claude Code or Codex on each machine that will do the work**, and handle the sign-ins when it asks.\n\n*YMMV!* 🤞\n\n```\n# T3 Code constellation: minimal setup\n\nGoal: workers run the agents and stay awake while they work.\nEvery other device is a remote control. Use ChatGPT and Claude\nsubscriptions, never API keys, for the agents in this setup.\n\nRules for the agent running this:\n- Hand every sign-in, approval code, and sudo command to the human.\n- Use the nightly channel everywhere, and update clients and servers together.\n\n## 1. Each worker: an Apple Silicon Mac or Linux box\nAgent:\n- Install Codex CLI, Claude Code, and GitHub CLI 2.81+ if missing.\n- `curl -fsSL https://t3.codes/install.sh | T3CODE_CHANNEL=nightly sh`\n- `t3 connect --headless`, then `t3 service install` if connect didn't.\n\nHuman:\n- `claude auth login` and `gh auth login`.\n- `codex login` at the machine, or later from a client:\n  Settings → Providers → Connect with ChatGPT.\n- Approve the link and code that `t3 connect` prints.\n- Mac only: stay logged in, and disable sleep with `sudo pmset -a sleep 0`.\n\nDone when:\n- `sh -lc 'command -v codex claude gh'` prints three paths.\n- `t3 service status` shows the service is installed and running.\n- `t3 connect status` shows `Environment link: provisioned`.\n  This checks saved setup; confirm live access from a client below.\n\n## 2. Each client\n- Desktop: the newest nightly from github.com/pingdotgg/t3code/releases:\n  the .dmg on macOS, the AppImage or .deb on Linux. Sign in to T3 Connect.\n- Remote-only laptop: Settings → Connections → turn off Local environment.\n- Phone: the beta app, from the QR codes in Settings → General → Mobile app.\n  Store apps can't connect to nightly servers.\n- Two or more workers: optionally enable\n  Settings → Connections → Load balancing for new desktop/web threads.\n  Choose the environment manually on mobile.\n\nDone when: every worker appears in the desktop app and on the phone,\nand you can open a thread on each worker from a client.\n\n## 3. Working agreement for agents inside T3\n- One task per thread, in a new worktree on a worker.\n- Open a PR once the checks you can run pass.\n- Get a second opinion from the other subscription: call\n  `orchestrator_capabilities`, then `delegate_task` with role `review`\n  on a different provider. Give it the PR's head SHA. It must not edit.\n- Fix real findings for at most two rounds, then stop and report.\n- Never merge. The human tests and merges.\n```\n\n## 🎰 Bonus\n\nWith all of my agents running on different machines, I’m like a clown spinning plates. I need to check in on more and more threads and keep them running (*the last frontier of humanity in software development, I’m afraid*).\n\nI subscribe to the theory that blindly copying and reusing skills can be a bit dangerous, or at least make your system feel weird to use.\n\nSo, I’m building my own skills, remixed from the ones [Matt Pocock and Poteto](https://x.com/mattpocockuk/status/2105239236178018636) have been sharing. My [bounds-skills](https://github.com/boundsj/bounds-skills) (WIP) bundle adapts planning, debugging, review, retro, and verification workflows to how I actually work:\n\n- 🧭 `bounds-mode` : one entry point that picks the right workflow and carries the task through its checks.\n- 📝 `bounds-plan` : turns an idea into scope, behavior to keep, and acceptance checks.\n- 🐛 `bounds-debug` : reproduces the bug, finds the cause, and proves the fix.\n- 🔍 `bounds-review` : reviews a pinned change and separates real findings from preferences.\n- 🔁 `bounds-retro` : spots repeated corrections and wasted effort, then proposes structural fixes.\n- 📚 `bounds-agent-docs` : keeps agent docs short, accurate, and loaded only when needed.\n- 🧪 `bounds-create-verification` : builds and runs a project’s own launch-to-cleanup test recipe, with evidence.\n- 🔧 `bounds-maintain-verification` : keeps that recipe current and tells doc drift from real bugs.\n\n**The goal is to keep my agents cooking when I’m not at home:** agents on the Mac mini and my big laptop…\n\n1. sort scope and write plans,\n2. implement and review, and\n3. do a full round of QA with visual evidence way before anybody needs to call me.\n\nBut, I’m on-call on T3 mobile when needed. 📟", "url": "https://wpnews.pro/news/environment-update-a-t3-code-tokenmaxing-constellation", "canonical_source": "https://www.bounds.dev/posts/environment-update-a-t3-code-tokenmaxing-constellation/", "published_at": "2026-10-05 02:00:00+00:00", "updated_at": "2026-10-05 03:41:22.909439+00:00", "lang": "en", "topics": ["ai-agents", "developer-tools", "ai-tools"], "entities": ["T3 Code", "Codex Desktop App", "Claude Code", "ChatGPT", "Claude", "pingdotgg", "M1 Mac mini", "M3 Max"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/environment-update-a-t3-code-tokenmaxing-constellation", "markdown": "https://wpnews.pro/news/environment-update-a-t3-code-tokenmaxing-constellation.md", "text": "https://wpnews.pro/news/environment-update-a-t3-code-tokenmaxing-constellation.txt", "jsonld": "https://wpnews.pro/news/environment-update-a-t3-code-tokenmaxing-constellation.jsonld"}}