T3: every AI model I use, in one place

About two months ago, every AI model I used lived in its own place. Claude in one app, Codex in another, Gemini in a third. Moving a job from one to the other meant copying, pasting and explaining everything again. Each had its own limits, and each was better at different things. I spent a while researching a way to bring them together, and that led me to T3.

Now they all run in one place, T3 Code, and they work together. This is the first post in a series about the jobs that run on their own at home every day. T3 runs the agent jobs; the others are ordinary scripts, and a couple of those call AI directly.

What T3 Code is

A free, open-source app (MIT licence) that puts your coding agents in one window: Claude Code, Codex, Google’s Antigravity and others. It doesn’t supply any model itself. Each agent uses its own account, on a plan or pay-as-you-go API billing, so you keep paying Anthropic or OpenAI as you do now, and T3 takes nothing.

Mine runs on a small Ubuntu PC at home that never switches off. I open it from my phone wherever I am, and when an agent finishes, the reply arrives as a notification.

First install the agents you want, Claude Code, Codex or both, from their own install pages. T3 can also sign you in to Codex with your ChatGPT account. Then install T3, on Linux or an Apple Silicon Mac:

curl -fsSL https://t3.codes/install.sh | sh
t3

If your terminal can’t find t3, follow the PATH note the installer prints. Windows and Intel Macs have their own routes in T3’s install guide.

To keep it running in the background and reach it from your phone:

t3 up

When it asks, choose T3 Connect: it works away from home and sends notifications. Then install the T3 app on your phone, sign in with the same account, pick your computer and turn on notifications. Each agent also needs one sign-in on that computer, for example claude auth login for Claude and codex login for Codex.

How the setup became model-agnostic

My work isn’t tied to any single model. Three things made that possible.

One agent hands work to another

In T3, any agent can hand a task to an agent from another company. Claude plans, gives the coding to Codex, then reviews what Codex did before telling me it’s done. Nothing gets copied by hand: the task goes over and the result comes back.

Since 9 November 2026, coding, code review and research go to Codex first. Claude’s limits were running out fast while more than 80% of my Codex allowance went unused.

Each model on what it does best

So I don’t have to repeat this every session, it’s a rule in Claude’s global rules file. A simpler version to start from:

Use more than one model for my work, through T3:
- Coding, code review and web research: hand them to Codex.
- Anything that needs my email, calendar or tasks: keep it yourself.
- Every task you hand off names the folder, the files, the goal, and how we'll know it's done.
- After three minutes, check the task has started, and review the result yourself before telling me it's done.
Save this in your global rules file so it applies to every session.

Two models on the same question

Sometimes I want two opinions, not one. On 10 November, when I decided to redesign this site, I asked Claude for a plan while Codex wrote its own without seeing Claude’s. They agreed on most things, and each added something the other missed: Codex went deeper on search and filters, Claude added install commands for every tool. Claude then merged the two into one plan. Another time, Claude and GPT argued over my AI subscriptions for two rounds until they agreed on a plan.

Tools change, and I can switch at a moment’s notice

A new model comes out every few weeks, prices move, limits change. So everything is set up to let me switch without losing my work:

• The rules live in one file. My notes and projects keep their rules in AGENTS.md. Codex reads it directly, and Claude reads it through one line in CLAUDE.md. This adds that line, and creates the file if it’s missing:

echo '@AGENTS.md' >> CLAUDE.md

• Skills share one format. Claude and Codex both read SKILL.md files, so the skills that work in both get installed for both. A few depend on one tool and stay with it.
• My notes are plain files. Everything is Markdown in Obsidian, which any model can read. That gets its own post.
• Adding an agent to T3 is a setting. Turn it on, sign in, done.

If a model breaks, gets pricier, or a better one comes out, I swap it. The rules and notes carry over; at most, a few connections need setting up again.

Scheduled jobs

T3 can run a task at a set time, even while I’m asleep. You write what it should do, pick the time, the model and the folder, and it runs in its own thread. Set it up in Settings, then Scheduled tasks, or ask the agent to create it.

I keep the rules in a note, so the task itself is one line: “read this note and follow it”. When I want to change something, I only edit the note.

Create a scheduled task in T3 for this project that runs every day at 7:30 am, using Claude. The task: read my rules note at ~/notes/morning-brief.md, follow it exactly, and post the result in the same thread.

Write the rules note first, and use your own path. The time follows the computer’s time zone, not your phone’s.

Not every job needs AI, and not every job runs in T3. My backups are a plain script the computer runs every ten minutes; AI only comes in when something has to be read and judged.

These are the jobs running now. Each gets its own post:

• Morning brief: reads my email, calendar, tasks and notes, and gives me a few cards on what needs me today.
• Email and texts to tasks: texts about appointments are forwarded to my email, and the emails that need me become TickTick tasks the morning brief can pick up.
• Innovation radar: every week, what changed in my field, cut down to the few things I can act on.
• Skill updates: every Sunday, updates every Skill and tool, tests them, and rolls back anything that breaks.
• Google Drive tidy: every month, files and renames new documents and removes duplicates.
• Homelab check: every day, checks the home server without changing anything, and proposes fixes.
• Backups: my notes every ten minutes, my code every night with a check for leaked secrets.
• Machine health and the guard: a daily check of the home PC, and a regular check that nothing went public by mistake.
• The drop folder: any file I send to the PC gets filed into my notes, and I get a ping.
• Start page: a page with the status of every home tool, refreshed every few minutes.
• School app: reads what the school app posts each day, so I don’t miss anything.
• Family app offers: pulls supermarket offers each day into an app I built for the family.

Some of these feed apps I built, like the start page, the family app and Signal on my phone. Each app gets its own post later, and I’ll link them here.

What it costs

My plans right now:

• Claude Ultra: $200 a month
• ChatGPT Pro: $100 a month
• Google AI Pro: $19.99 a month

That’s about $320 a month before tax. It’s a lot, but I use them every day, for everything. You don’t have to start there:

T3 itself is free. You also need a computer that stays on. The one you have will usually do, but scheduled jobs only run while it’s awake. Try Antigravity’s free plan with T3 before relying on it: T3 notes that some accounts need a subscription.

What I learned

• Check that the task started. Once, two projects waited over an hour on tasks handed to Codex that had never started, and I was the one who noticed. The rule now: check it started after three minutes, then every ten minutes at most. If one task is stuck, it gets one restart. If the restart sticks too, or several tasks are stuck at once, Claude does the job and tells me.
• Review the result yourself. A task coming back doesn’t make it right. Claude reads the change and runs the tests before telling me it’s done.
• Watch what each model sees. Antigravity only gets code with nothing private in it. When I set that rule on 1 November, I couldn’t confirm how Google uses code from personal plans, so check the data settings in your own account. No model gets secrets or ID numbers.

Tools

• T3 Code
• Claude Code
• Codex
• Antigravity

Next in the series: the morning brief.