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.
Past that, two choices are yours, and they come down to how you like to work: how you talk to it (at a desk when you choose, or in a chat app that comes to you), and which machine it drives for sites you're signed in to. Both are laid out below, with what I do. 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.
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 most 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.
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.
ssh to the VM by its tailnet name using a key, with no password prompt.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 backups. Proxmox Backup Server installed beside Proxmox takes the half-hourly snapshots, and a second small box running Proxmox Backup Server pulls a copy. Proxmox can only push its backups, so the pull is what leaves the first box with no credentials for the second.
A gaming PC is a different thing: a daily machine for you, on its own budget. I'd keep it out of this.
If you'd rather work at a desk, you open T3 when you choose, on your PC or phone, and give it jobs. For the proactive part, a scheduled task in T3 runs a check on a timer: it reads your mail and messages and writes up what needs you. This is the simpler of the two, because it's one system. Set each check to "One thread, fresh session each run", so its reports collect in one place and each run starts clean. T3's phone app can push notifications through T3 Connect. I haven't tested whether a scheduled check's report triggers one.
If you'd rather it came to you in a chat app, Hermes is what I run for this. It's an open-source assistant that lives on its own VM and talks to you in a chat app on your phone (it supports several, Telegram among them). Its scheduled jobs can send you a morning briefing or a summary of your messages, and it keeps memories of what matters to you. I've used it with my ChatGPT plan. The cost is a second system to look after, with its own jobs, memory and updates, so the managing applies twice. Give it an SSH key to the hub so its jobs can use the same mail and message readers. I still send work that needs a signed-in screen or a long build to T3.
Either way, have every check report, for each source, what period it covered, how many new items it looked at, and whether that source actually synced. Zero new messages can be a quiet day or a reader that stopped syncing days ago, and only the sync status tells them apart. Give it a short file of what you care about, which it reads every run and you correct when it gets something wrong. That file is what makes it good, more than the model.
Your mail is read straight from the hub, with nothing switched on at your end: gog for Gmail once you've authorised it, or IMAP if your provider offers it. Some providers need an extra piece or a paid plan for IMAP, so have your agent check yours.
WhatsApp has three routes:
wacli and ZapFast are unofficial clients, and WhatsApp bans accounts it decides are automated, so both carry some risk to your account. For SMS, Google Messages for web works the same way as WhatsApp Web: left open on the driven machine, paired with your phone, which needs Google Messages and a data connection.
Banking, the ATO and any booking that needs an account need you signed in, so the agent drives a real machine with screenshots and real mouse and keyboard after you sign in. Three ways to try it:
On whichever machine it drives, an SSH session can't see the desktop on its own, so the agent needs a small helper that starts when you log in and lets it take screenshots and click inside your session. That helper is the piece you build rather than install. Xav's quick scheduled-task route in Part 3 is enough for a first version. Mine is about 1,500 lines now, grown over four months of use. 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. Check: with your own session on that machine closed or the lid where it lives, the agent takes a screenshot and clicks something harmless.
| You want | How it happens here |
|---|---|
| Taxes, ATO, banking | You sign in on the machine the agent drives. It takes over from there, in a real browser. |
| Flights and bookings | Research and comparison from the hub. A booking that needs your account happens on the signed-in machine. |
| Finding things you'd like | Mostly research with no login. It gets better as it keeps notes on your taste. |
| Monitoring and flagging issues | A scheduled check, as a T3 task or a Hermes job, reads your mail and messages and tells you what needs you and what it read. |
| Messages | Mail through gog or IMAP from the hub. WhatsApp through wacli, ZapFast or WhatsApp Web, and SMS through Google Messages for web, as above. |
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.
rm protects nothing. Plenty of other commands delete things. I run with permissions off and accept that, which is at least honest about the risk.Build spec, in the order above. Steps 1 to 5 are common. Ask before building past them.
Host: the GMKtec G10 mini PC, Proxmox VE, wired ethernet,
headless. BIOS: power on after AC loss; hub VM set to start
at boot. Check both return after a power cycle.
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 the hub VM, Remy's phone and PC, and any machine
you drive. The phone reaches T3 at its tailnet address; T3
Connect only if that proves insufficient. No ports opened to
the internet.
Secrets: env files read at runtime, never pasted into chat or
printed in output.
Roles: Claude manages; hand reviews and bounded legwork to Codex.
Before building past step 5, ask Remy these one at a time and keep
his answers in a file you both read:
1. Desk or chat: does he open T3 when he chooses, or want an
assistant that messages him?
2. Which machine you drive: his PC, a laptop of its own, or a
desktop VM? He said he doesn't want it "able to control
anything except the PC". Which machine did he mean?
3. Interrupt him, or brief him at set times? Which hours are quiet?
4. Whose messages always matter?
5. Which three chores does he hand over first?
Desk: each recurring check is a T3 scheduled task set to "One
thread, fresh session each run".
Chat: Hermes Agent (github.com/NousResearch/hermes-agent) on its
own VM, connected to the chat app he picks, with an SSH key
to the hub so its jobs run the hub's readers. Check: a
scheduled Hermes job reads a known item through the chosen
reader and delivers it to his chat.
Checks: per source, report the period covered, new items examined
and whether the source synced; name any source that failed
or couldn't be verified. Read a short file of what he cares
about every run and update it from his corrections.
Mail: find his provider. gog for Gmail once authorised; IMAP if
the provider supports it.
WhatsApp: the route he picks. wacli on the hub (report its message
count each run; upgrade when WhatsApp rejects the version),
ZapFast's archive read-only on the machine it runs on (a
reader you build; report whether ZapFast is running and the
newest archived message), or WhatsApp Web on the machine
you drive.
SMS: if he wants it, Google Messages for web on the machine you
drive, paired with his phone.
Driven: his PC or the laptop. SSH key auth from the hub (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. Then run commands in his
session: start with Xav's Part 3 scheduled task (/IT), then
a small helper started at logon ("desktop agent") for
screenshots, clicks and keys. The machine must be on, signed
in and unlocked. Check: with the lid or session as it
normally sits, take a screenshot and click something
harmless. Remy signs in to sites; you take over after.
Desk VM: a second VM with a desktop and a browser, and the same kind
of helper inside it. Check the screenshot and click with
Remy's Proxmox console closed. Test his bank and the ATO in
it before relying on it. Keep it light: 16 GB is shared
with the hub.
Later: Proxmox API token for creating VMs. Proxmox Backup Server
beside Proxmox, half-hourly. A second box running PBS
pulls a copy with a read-only token, so the first holds
no credentials for it.