My agent setup, November 2026
A follow-up to my Claude Projects post: local machines, threads, and cross-provider agent comms with T3 Code.

A month ago I wrote about moving my personal workspace into a Claude Project. By the time that post went up, I'd found Remote Control, which lets a project reach my own computer. I was still passing jobs between Claude and ChatGPT by hand, through a TASKS.md file.
During October I gave the project machines that stay awake, and connected it to T3 Code so it could hand work to other agents itself. This is the setup as of November.
What a cloud thread can't do
A project thread runs in the cloud. It can read the web and push to my repo, but unlike ChatGPT, Anthropic doesn't have a cloud browser that keeps my logins safely between sessions. Even ChatGPT's can't handle more than one profile.
Sites are also starting to block traffic that comes from cloud browsers.
Remote Control covers both. The thread starts a Claude Code session on one of my machines, that session works with the machine's own files, tools and browser from my home connection, and the result comes back to the thread.
Choosing a machine
My Mac was the first target. It has Claude in Chrome with my own sign-ins, and computer use. It's also my personal laptop, and it sleeps when I close the lid, so a job I start from my phone at night can't count on it.
So I added two machines that never sleep.
A Raspberry Pi came first. It handles pings and light jobs, like fetching pages from sites that turn cloud servers away, and it drives my Pixel over wireless ADB.
An old laptop does the coding. It also runs two browsers just for agents, one for personal accounts and one for freelance work, each signed in to a few sites I'm comfortable letting an agent use headlessly.
Not being the messenger
The TASKS.md handoff worked, but I was the relay. I told ChatGPT to check the repo, then told Claude that ChatGPT had finished.
Until recently I only used T3 Code at work, to drive several coding agents from one control pane. I could start a thread on a project folder in Claude Code, another in Codex, and so on. Over the last few months it has grown into something much bigger.
First it learned to reach other machines. T3 Connect makes a machine running T3 Code reachable from my other devices without port forwarding, and linked environments let agents on one machine work on another. My old laptop is the hub, with the Mac and the Pi linked to it. I check in from the desktop app on the Mac or the phone app on my Pixel.
Then came the orchestrator. A single chat, much like a Claude Project's coordinator, can hand tasks to child agents on any provider and model, across machines, projects and threads. It went stable in T3 Code 0.1.0 after running in nightly builds since early October.
The T3 Code MCP gives outside agents a way into all of that. An agent can read projects and threads, start and message threads, and see which providers and models are available.
That's how my Claude Project uses it. A thread can start Codex on the old laptop, wait for it and read the answer, without me carrying anything across. It also means I can ask GPT models, or Gemini through Copilot, for reviews.
Why not just T3 Code?
T3 Code is great for development work. I also like its cross-project communication, which is lacklustre in Projects, and that I can start a chat without a project.
Still, I get a better experience from Claude Projects, because it feels designed for orchestration first, not just code.
Projects has plenty of UI bugs. Threads don't show links to the sessions they start on my machines unless I ask, approval requests sometimes drop, and there's no "always allow" for folder access.
Even so, I prefer its layout. Each thread shows up as a card in the coordinator's chat, and opening one puts it in a side panel, so I can quickly skim through them.
T3 Code's orchestrator is a single chat. Children report back as links, and the lineage pane lists them, but opening one replaces the whole chat. To get back, I go up through the lineage pane, which gets annoying.
I also don't have to set up or maintain the machines cloud threads run on. That's because I'm lazy. T3 might one day spin up machines for me with all my providers signed in, but it doesn't today.
Where this leaves me
I really want to try Claude Duet. Anthropic is building it as an always-on agent that sits above projects and talks to their threads. It's out, but US only for now.
At the end of October Duet started working across Claude Projects. It still can't start work on my machines. Karan Ahuja from the Duet team said on X that this is their highest priority and should land early next week.

If T3 Code and Duet keep improving at this pace, they might replace Projects for me.
Next on my list is Tailscale, so I can check on the agent browsers away from home without putting them on the open internet. I'll redraw this when it's in.