For Remy ยท 21 Nov 2026

Your agent box, and Xav's guide

Written for you and readable by your agent. Hand it the link.

Put Proxmox on the 16 GB box, make one Linux VM your agent's home, and reach it from anywhere over Tailscale. Claude Code and Codex live in that VM, T3 Code lets you talk to them from your phone or PC, and the work keeps running when you close the device. You keep using Android and Windows. The models run on your subscriptions, so the box only needs to be reliable, not powerful. For sites you're signed in to, your agent drives your own Windows PC over the tailnet. Xav's guide is right that a real signed-in browser gets blocked far less than a bot browser, which is the problem you're hitting with Dot. Most of what it builds beyond that solves problems this setup doesn't have.

What you're after, and what it asks of you

You want an EA that's proactive and handles everything: life admin like taxes and flights, research, finding things you'd like (products, games, books), monitoring your email and messages, and bringing issues to your attention. This setup can reach all of it.

The part people underestimate is that it asks you to manage it. It isn't a magic box. When something goes wrong, tell it what problem you're having and ask for ideas to fix it. When it does something you don't understand, ask why, and how it can avoid that in future. Check in, ask how it's going and what's blocking it, and offer resources when it needs them. You're never fully hands off, but the time you put into understanding how it works is what makes it better.

How I run mine

How you'd do something similar

Roughly in this order, checking each step works before the next. You do steps 1 to 4 by hand. It's a bit hairy the first time, and a good way to learn how the pieces fit. From step 5 your agent can do most of the work with you.

  1. Proxmox on the box, plugged into ethernet, no monitor once it's installed. Check: the Proxmox web page opens from your PC.
  2. One Linux VM as the hub. Check: it boots and reaches the internet.
  3. Tailscale on the VM, your PC and your phone. Check: from your PC, ssh to the VM by its tailnet name using a key, with no password prompt.
  4. Claude Code and Codex on the VM, each signed in to its plan. Check: each finishes a small task.
  5. T3 Code on the VM. Check: pair your phone with the server's tailnet address, start a thread, lock the phone, and it has made progress when you come back.
  6. Your Windows PC as the signed-in desktop: Tailscale, OpenSSH Server, the VM's key, and a small helper that starts when you log in so the agent can take screenshots and click inside your session. An SSH session can't see your desktop on its own. Xav's Windows notes are the reference for your agent here: Part 2's OpenSSH steps and admin key file, display scaling and the RDP warning, and Part 3's desktop agent. Some of his Part 2 settings are meant for a dedicated machine, so pick what suits your own PC. The PC has to be on and signed in for these jobs, and the agent moves your real mouse. Check: the agent takes a screenshot of your desktop and clicks something harmless.
  7. Later, once you trust it: give the agent the Proxmox API so it can make its own VMs.

Plans. Start with a Claude Max subscription as the manager. Keep ChatGPT Pro for now, watch how much Codex actually uses it, and consider dropping it to a cheaper tier, but keep the account. A second model family reviewing the first one's work earns its place.

After a couple of months, if it's earning its keep, add a second small box that takes half-hourly snapshots of your VMs. Give the first box no credentials for the second.

Your wish list, mapped

You wantHow it happens here
Taxes, ATO, bankingYou sign in on your PC. The agent drives from there, in your real browser.
Flights and bookingsResearch and comparison from the VM. The booking itself happens on your signed-in PC.
Finding things you'd likeMostly research with no login. It gets better as it keeps notes on your taste.
Monitoring and flagging issuesA job on a timer on the VM, which your agent can set up, that checks the mail and messages you point it at and tells you in T3 what needs attention.
MessagesWhatever you can open on your PC, it can read there.

Xav's guide: good and bad

The guide is clearly written, and some of it is hard-won. It's right that a real, signed-in browser gets challenged far less than a bot browser. It's right to read the page's structure rather than screenshots wherever possible, its account of why an SSH session can't see the desktop is accurate, and the lesson that Claude grades its own work generously is a good one. Its rule that the agent drafts any message to a person and waits for your OK is sensible, and the password vault, which types a saved password only into the matching site without Claude ever seeing it, is a good design.

For your agent

Build spec, in the order above:

Host:     the mini PC, Proxmox VE, wired ethernet, headless.
Hub VM:   one Linux VM; Claude Code + Codex CLI installed and signed in;
          T3 Code server (github.com/pingdotgg/t3code) bound to the
          tailnet address only.
Network:  Tailscale on hub VM, Remy's Windows PC and phone. No ports
          opened to the internet.
Access:   SSH key auth from the hub to the Windows PC (OpenSSH Server).
          For an admin account the hub's public key goes in
          C:\ProgramData\ssh\administrators_authorized_keys, with
          access limited to SYSTEM and Administrators. Test key login
          on its own before desktop control.
Desktop:  an SSH service can't see Remy's desktop. Run a small helper
          started at logon in his session (Xav's guide, Part 3,
          "desktop agent") for screenshots, clicks and keys. The PC
          must be on and signed in. Remy signs in to sites; you take
          over after.
Secrets:  env files read at runtime, never pasted into chat or
          printed in output.
Roles:    Claude manages; hand reviews and bounded legwork to Codex.
Later:    Proxmox API token for creating VMs. Half-hourly snapshots to
          a second box that pulls them, so the first holds no
          credentials for it.