This is my actual setup, not a universal recipe. Tweak the device list, providers, permissions, skills and MCPs for your own work. IMPORTANT TAILSCALE NOTE My topology assumes a Linux server connected to the same Tailnet as my other devices. I use Tailscale because I have more than three machines and want one private network for SSH and services. T3 Code does not require Tailscale. T3 Connect works without it and is enough for most people who only want to reach their T3 environment remotely. Do not add a Linux server or Tailscale just to copy me. THE ARCHITECTURE iPhone or another computer → T3 Connect, with no VPN required or → Tailscale, as an optional private device network → T3 Code on the host Mac → Codex, Claude or Grok provider CLI → project files, terminal, Git, preview, skills, MCP tools and local subagents WHAT I RUN • Apple M5 Max MacBook Pro with 128 GB RAM • T3 Code 0.0.33 • Codex, Claude and Grok enabled • Cursor Agent installed but disabled in T3 • OpenCode disabled • Codex: GPT-5.6 Sol, xhigh, priority • Claude project default: Fable 5 • Grok 4.6, xhigh, yolo off • Node 26, pnpm, Bun, Git, gh, Docker, uv and 1Password CLI T3 CONNECT AND TAILSCALE T3 Connect links the local T3 environment to a T3 account and uses a managed cloudflared tunnel with authenticated HTTPS/WSS. It keeps the repository, terminal and running process on the host. Tailscale joins devices to a private Tailnet. I use it for SSH and services across my Mac, iPhone, Linux server and other machines. Tailscale Serve is optional. Funnel is public and is not part of this setup. Current T3 install guide: https://github.com/pingdotgg/t3code/blob/main/docs/user/install.md T3 Connect internals: https://github.com/pingdotgg/t3code/blob/main/docs/internals/t3-connect.md Optional Tailnet HTTPS: npx t3 serve --tailscale-serve Tailscale for macOS:

If you keep handing work to coding agents and they keep giving you back something broken, the fastest fix is usually to give the agent something to read. Screenshots confuse agents on anything complex, because there is a server and a database and a front end, and the agent ends up guessing at which one failed. So add a logging system. Log the timing from when a user clicks to when the page actually finishes rendering, not just the API call. Log the round trips between your database and your front end. Then build a way to export a log bundle, so you can freeze the exact moment the bug happened and hand that state to the agent. When your agents are testing while you are asleep, they can export that same bundle. Now they have the artifact they need to actually verify a fix instead of guessing at one. This is what we were doing at Apple long before AI. You would basically just grep for your error keywords. The agents have seen thousands of examples of this in their training, so once you have it, they already know what to do with it. This came out of the first AI Builder Room this week. Try it and tell me what happens.

Sorry it took me so long and ive been a little quiet in the live streams / discord i been super busy joining so i can assist people who need it @Ray Fernando keep up the amazing work.