Build Your Own AI Agent: A Private, Always-On Assistant You Actually Own

Time ≈ 1.5-2 hours Server cost ≈ $24–29 / month AI your choice (subscription or API) Software cost $0 (Hermes is free)

Each step is tagged: PASTEcopy my text into the console YOUonly you can do it (sign-ins, payment) CHECKconfirm it worked before moving on

💡 Tip: click any grey command to copy it, the whole thing, ready to paste.

What makes this powerful

It's an agent, not a chatbot. Chatting with an AI is limited: you ask, it answers, you still do the work. Hermes uses tools, writes and runs code, browses, and finishes multi-step jobs on its own, around the clock.

It continuously learns and maintains itself. It keeps a persistent memory and builds skills as you use it, restarts itself, and runs scheduled jobs, so it gets more useful over time instead of forgetting every chat.

Free, and any AI you choose. Hermes itself is free and open-source. You can switch to the best or newest model the day it launches (Claude, Grok, GPT, Gemini, open-source), with no lock-in.

Reachable from your Mac and your phone, from anywhere on the planet. This is the system that keeps you on the frontier of AI progress: one place that plugs into whatever the best model is, learns as you use it, and works while you sleep.

Read this first

You cannot break anything. This new computer is completely separate from your websites and Cloudflare. Worst case, we delete it and start over in 10 minutes.

No password or card ever leaves your own screen. Every sign-in and payment happens in your browser, on your machine. That is by design.

Stuck at any point? Every step has a "what you should see" check and, where things commonly go wrong, a troubleshooting note. Tick the boxes as you go, this page remembers your progress.

What this is, and why I bothered

I set this up for myself. Here is the reasoning in plain terms, one box at a time.

What is Hermes?

Hermes is a harnessiA harness is the software wrapper around an AI model that gives it memory, tools (files, terminal, web, messaging), scheduling, and the ability to act on its own, instead of only replying in a chat.: software that wraps an AI model and turns it into an agent that can remember, use tools, run on a schedule, and act on its own. Think of it like OpenClaw, but much easier to set up and to live with.

The model (Claude, Grok, GPT, and so on) is the brain. Hermes is the body, hands, memory, and calendar built around it.

Why run an agent like this at all?

Most people use AI by opening a chat and typing. That is powerful, but you are the loop: you ask, it answers, you ask again, and the moment you close the tab nothing happens.

A harness flips that around. It runs 24/7, remembers everything, works on tasks while you sleep, checks things on a schedule, and messages you when something needs attention. You stop being the one clicking and start being the one delegating.

It also lets you build an org chart: a "chief of staff" agent you talk to, which hands work to specialist agents (research, coding, ops) and reports back to you. That is the real unlock, and it puts you in the tiny fraction of people who don't just use AI, but have AI genuinely working for them.

Why not just keep using Claude Code, Codex, or a Grok chat? Those are single tools you drive by hand, one conversation at a time. Hermes actually uses those tools, but adds the memory, autonomy, scheduling, messaging, and multi-agent structure a plain chat window will never have.

Why put it on the cloud, and not just my laptop?

An agent is only useful if it is always on, and your laptop can't be. The lid closes, the machine sleeps, and the agent sleeps with it. You would also have to leave it pinned open with the fans running, unable to use the machine normally, and everything would stop the moment you travel or restart.

A cheap always-on cloud computer fixes all of that: it never sleeps, survives reboots, keeps working while your laptop is shut, and lives at a stable address on the internet.

Could I use a spare Mac at home instead? Yes, if you own the hardware, can genuinely leave it powered on 24/7, and are comfortable securing it. For most people a ~$24/month cloud box is simpler, more reliable, and kept safely separate from the computer they actually use every day.

First, three quick decisions

The defaults below are what I used. Open a box only if you want the reasoning or a different option.

Why self-host this instead of a managed cloud (like Nous Cloud)?

You own everything. The agent, its memory, your sign-in tokens, and every conversation live on a machine only you control. A managed host runs all of that on their servers, so you are trusting them with your data and your account access.

It reuses subscriptions you already pay for. If you already pay for a Claude, Grok, or ChatGPT plan, it works here at no extra per-use cost. Managed hosts usually charge their own way (their plan, or per token) on top.

No lock-in. You can back it up, move it, or rebuild it anywhere. This guide even sets up free backups.

The trade-off: you do a one-time setup (this guide) and you own the upkeep. If you would rather pay for zero-setup convenience and accept less privacy and control, a managed cloud is the faster route. This guide is the private, in-your-hands route.

Which cloud computer (VPS) should I use?

Any provider that rents an Ubuntu VPSiA VPS (Virtual Private Server) is a slice of a real server you rent by the month. It behaves like an always-on computer in a data centre that only you can log into. works, so this matters less than it first looks. Here is how I actually thought about it.

The thing to realise: this is a machine you set up once and then leave running for months. It quietly holds two things you cannot easily recreate, your signed-in sessions to the AI services and the agent's memory, which grows the longer it runs. So I weighted boring reliability, one-click snapshots, painless support, and simple billing far above shaving a few dollars off the monthly price. With that lens:

DigitalOcean: My pick, and what this guide's screens match. Its console is the friendliest for non-experts, its snapshots and backups genuinely "just work," and billing is simple and US-based. It costs a little more than the cheapest option, and for a machine holding my sign-ins I happily pay it. The safest default, especially in the US.
Hetzner: The value pick: roughly half the price for the same RAM and CPU, with data centres in Germany, Finland, and now the US too. Genuinely great hardware for the money. Three honest trade-offs, none of them fatal: support is leaner, signup sometimes adds an ID-verification step, and its data-centre addresses are among the most likely to trip "are you a bot?" checks when your agent browses the web. If price is your priority it is a fine choice, and the steps here are nearly identical, only the signup and "create server" screens differ.
Others: Vultr, Linode, and similar work fine too. The one rule that matters: pick one with a data centre near you.

Which region? Pick the data centre closest to where you sit, for two reasons: the control panel feels snappier, and your sign-ins stay in a consistent part of the world (a login that suddenly appears from a far-off country can trip an account's own fraud checks). One thing that does not matter here: where your website's visitors are. This box serves only you and your agent, never your site's audience, so ignore your traffic map and think only about where you are. Europe is the answer only if that is where you are.

Size: 4 GB RAM / 2 CPUs is plenty to start, because the AI itself runs on the model providers' servers, not on this box. Roughly $24/month on DigitalOcean, a little less on Hetzner. If you later run many agents in parallel, you can resize up in two minutes.

Which AI provider(s) can I use?

Hermes is provider-neutraliHermes is the harness, not the AI. It can drive almost any model, and you can plug in several and switch per task.: it drives almost any model, and you can add several and switch per task. The honest landscape (this space moves fast, and more of these are adding flat-rate subscriptions over time, so re-check before you commit):

Claude (Anthropic): Subscription (Claude Max) or API. Widely rated the strongest all-round coder.
Grok (xAI): Subscription (SuperGrok) or API.
GPT (OpenAI): Subscription (ChatGPT / Codex) or API.
Gemini (Google): Mostly API today.
Open-source: Llama, Qwen, DeepSeek, and Kimi (Moonshot, whose next open release is expected around late July 2026). Run locally with Ollama for full privacy and no per-use cost, or cheaply via a host. Good local speed needs a bigger machine.
OpenRouter: One API key for hundreds of models, handy for trying several.

How to choose: a subscription is best if you already pay for one (no extra per-use charge); an API key is pay-as-you-go and works with everything; local models give maximum privacy and zero per-use cost if you host them. You set this at the "choose a provider" step (or later with hermes model), and you can add more than one.

What I used: I already paid for Claude and Grok subscriptions, so I wired up both to avoid new per-use bills. If you are starting fresh, use whatever you already have, or an API key for a model you like. Nothing in the steps below depends on my specific choice.

How the pieces fit together

In one sentence: your agent lives on a small rented computer that is always on, you reach it privately from your Mac or phone, and the heavy AI thinking happens on the model companies' servers.

You

💻 Mac app · 📱 iPhone (HermexiHermex is a free (community-made) iPhone app for chatting with and watching your agent. It connects privately to your cloud computer over Tailscale. Optional, set up in Part 8.)

How you talk to and watch the agent. Close them anytime; nothing on the agent stops.

☁️ Cloud computer always on

Hermes, the agent

Runs 24/7 with its coding helpers Claude Code and Grok Build. This is the box you build below.

🧠 The cloud computer sends requests out to Anthropic (Claude) and xAI (Grok), where the AI actually runs, billed to your subscriptions.

Open its screen & prepare it

≈ 5 min

The droplet has no monitor, its "screen" is a black text window that opens right in your browser. Everything from here on is copy → paste → wait.

Click your droplet's name (hermes), then the Web Console button at the top of the page, next to the name (also under Actions → Connect). A terminal window opens ending in root@hermes:~#, that means "ready". If it shows login: instead, type root, Return, then your droplet password (typing stays invisible, normal).
PASTEFirst, start a tmuxitmux keeps your programs running on the droplet even if your browser or connection drops, you reconnect and they're exactly as you left them. session, a safety net so a dropped console never loses your place. Type this and press Return:
keep your work safe
```
tmux new -s setup
```
The screen clears to a fresh prompt with a green bar along the bottom, you're now inside tmux. Do everything from here in this window.

If the console ever disconnects: reopen it (same Web Console button), then type tmux attach -t setup to land exactly where you left off. If you weren't in tmux when it dropped, just run hermes setup to re-open the wizard, your finished sign-ins are saved, so you only redo the unfinished picks.

Stuck, can't type, or a command seems to do nothing? You're probably still inside a program. Press Ctrl + C to back out of most things; if a page of text is filling the screen and won't scroll, press q; to leave Claude Code type /exit. You can type commands again when you see root@hermes:~# on its own line.
PASTECopy Block A below (use its Copy button), click inside the terminal window, and paste with ⌘V. Press Return if it doesn't start by itself.
BLOCK A updates the system + turns on the firewall
```
# Block A, prepare the computer (takes 2–4 min)
export DEBIAN_FRONTEND=noninteractive
apt-get update && apt-get -y upgrade
apt-get -y install curl git tmux unzip fail2ban unattended-upgrades
ufw allow OpenSSH
ufw --force enable
fallocate -l 2G /swapfile && chmod 600 /swapfile && mkswap /swapfile && swapon /swapfile
echo '/swapfile none swap sw 0 0' >> /etc/fstab
```
What does Block A actually do? (safe to skip)

All standard first-setup commands, nothing destructive, and it only touches this new droplet, never your websites or Cloudflare. Line by line:

export …noninteractive
Tells installers not to pause and pop up blue question screens.

apt-get update && upgrade
Updates the machine's existing software to the latest security patches.

apt-get install …
Installs 6 tools: curl (downloads the next installers), git (needed by the coding agents), tmux (keeps programs running in the background), unzip, fail2ban (auto-blocks password-guessing attackers), unattended-upgrades (auto-installs future security updates).

ufw allow OpenSSH
Firewall: allow your own console/SSH connection in.

ufw --force enable
Turns the firewall on (blocks everything else).

fallocate … swapon
Creates a 2 GB "swap file", overflow memory on disk, so the 4 GB box won't crash if it briefly needs more.

echo … >> /etc/fstab
Makes that swap survive reboots. The only line that edits a system file, it just adds one line to the swap config.

A wall of text will scroll past, completely normal. If a pink/purple box ever appears asking about services, just press Return to accept.
CHECKWait until the scrolling stops and you see root@hermes:~# again on its own line. Last message should include Firewall is active.
Each time you open the console you'll see an "N updates can be applied" banner and an "ESM Apps / Ubuntu Pro" pitch, that's normal, not an error. Security updates install automatically (Block A set that up), so there's nothing to do; ignore the ESM/Ubuntu Pro upsell. To apply pending updates immediately anyway: apt-get update && apt-get -y upgrade.

Sign in to Claude & Grok

≈ 10 min

The one dance you'll repeat: the terminal window shows a link → you open that link in your normal Mac browser → sign in → copy the code it gives you → paste the code back into the terminal window. Your passwords never touch the droplet.

Is using my subscriptions this way allowed?

Broadly yes, you're using each vendor's own official tool and login, on your own account. xAI explicitly connected SuperGrok to Grok Build (official sign-in, no API key), and Claude Code is Anthropic's own CLI, running it headless on a server is normal, supported use.

What keeps it clean: it's your personal account (don't share or resell access), you use the official sign-in flows (these steps), and you stay within normal personal fair-use, a 24/7 agent can bump subscription rate limits, and industrial-scale automation could cross the line.

One nuance if you use Claude: Anthropic has been tightening how Claude Max subscriptions may be used by third-party agent tools, which is why this guide has Claude help only through its own first-party Claude Code CLI, rather than wiring a Claude login into Hermes as its model. Not legal advice, and terms change, so glance at your provider's current terms, but as built this is an intended way to use a subscription you already pay for.

YOUType claude and press Return. If it asks "Do you trust the files in this folder?" choose Yes, it's your own server. Pick any color theme (Return), then choose Log in with your Claude account (the subscription option, not "API key"). Select the long link it prints with your mouse, copy it (⌘C), open it in Safari/Chrome, sign in to Claude, then copy the code you're given and paste it back into the terminal window.
Expect: "Logged in" with your Claude account email. Then type /exit and press Return to leave Claude Code.
YOUOnly if you installed Grok Build: type grok login and press Return, same dance with your X / SuperGrok account. This logs in the Grok Build coding CLI; it's separate from the Grok brain sign-in you did in the wizard. Skip if you skipped Grok Build.
Seeing "Add credits to continue using the API · $0.00"? Grok Build is trying to bill the pay-per-token API instead of your subscription, which means the account you logged in with isn't on a consumer SuperGrok plan. Grok Build is included with personal SuperGrok / SuperGrok Heavy, not with Business / Team plans (those are console/API-based). Fix: grok login with an account that has SuperGrok (Heavy), and confirm with xAI which plan includes Grok Build before you pay, since it's account-specific. (Same reason a Business plan hides the personal usage meter, team usage lives in the admin/billing console.)
CHECKPaste grok --no-auto-update -p "Say ok.", Grok should answer with a little "ok". That proves your Heavy subscription is doing the work, no API billing.
CHECKHermes's own brain (xAI Grok) was already set in Part 3, nothing to do here. To swap it later, type hermes model anytime (Part 3 explains why Claude isn't the brain).

Turn Hermes on, permanently

≈ 5 min

Two pastes: give Hermes's control panel a login of its own, then register Hermes as a service so it runs 24/7 and restarts itself after any reboot.

YOUInvent a password for the Hermes control panel (different from the droplet one) and type it here, the box below customizes itself. Save it as "Hermes dashboard" in your password manager, you'll type it (with username admin) into the Mac app in Part 7. If it's ever lost, it's recoverable from ~/.hermes/.env on the droplet.
Dashboard password:Stays on this page only, it's just filled into Block E for you. This page never sends anything anywhere.

PASTE

BLOCK E control-panel login

mkdir -p ~/.hermes
cat >> ~/.hermes/.env <<'EOF'
HERMES_DASHBOARD_BASIC_AUTH_USERNAME=admin
HERMES_DASHBOARD_BASIC_AUTH_PASSWORD=PUT-A-PASSWORD-HERE
EOF
echo "HERMES_DASHBOARD_BASIC_AUTH_SECRET=$(openssl rand -base64 32)" >> ~/.hermes/.env
chmod 600 ~/.hermes/.env

PASTE

BLOCK F run Hermes as a 24/7 service

source ~/.bashrc 2>/dev/null
HERMES_BIN=$(command -v hermes)
cat > /etc/systemd/system/hermes.service <<EOF
[Unit]
Description=Hermes Agent
After=network-online.target

[Service]
ExecStart=$HERMES_BIN serve --host 0.0.0.0 --port 9119
Restart=always
User=root
Environment=HOME=/root
Environment="PATH=/root/.local/bin:/root/.hermes/bin:/usr/local/bin:/usr/bin:/bin"

[Install]
WantedBy=multi-user.target
EOF
systemctl daemon-reload && systemctl enable --now hermes

CHECKConfirm Hermes is running, paste this:
check Hermes is running
```
systemctl status hermes
```
Look for active (running), it shows up in green, exactly like that. Then press q to start typing again (that status view holds the screen until you press q). Hermes now runs around the clock; the firewalliA firewall blocks all incoming connections from the internet except the ones you allow, so the control panel can't be reached from outside. keeps its control panel invisible to the internet, Part 6 gives you a private way in.

Private line to your Mac

≈ 5 min

TailscaleiTailscale links just your own devices into one private, encrypted network, like putting your Mac and the droplet on the same home Wi-Fi, wherever they are. is a free app that creates a private tunnel between your Mac and the droplet, so the Hermes panel is reachable by you and no one else. (It coexists fine with your regular VPN.)

Wait, if I use the Mac app, why do I need Tailscale?

Tailscale isn't instead of the Mac app, it's what lets the app reach your droplet safely. The app has to connect to Hermes running on the droplet (an address + port 9119), and there are only two ways to allow that:

Open that port to the whole internet, then anyone can find your agent's control panel (guarded by just a password), and bots scan for exactly this. Risky.

Tailscale, a private tunnel only your Mac can use; the droplet's panel stays invisible to everyone else. Free, set up once, runs quietly. It's also what Hermes's own security guidance recommends.

So you still run everything through the Mac app, Tailscale is just the private road it drives on.

PASTEOn the droplet:
BLOCK G Tailscale on the droplet
```
curl -fsSL https://tailscale.com/install.sh | sh
ufw allow in on tailscale0
tailscale up
```
Expect: it prints a login link. Open it on your Mac and sign in (GitHub, Google, Apple, or email, any is fine), that creates your free Tailscale account. Remember which you pick; you'll use the same one on your Mac next. Let tailscale up finish after you approve in the browser, don't press Ctrl+C while it's waiting.

If tailscale ip -4 later says NeedsLogin or "no current Tailscale IPs": the login didn't finish. Just run tailscale up again, open the link, approve the hermes machine, and let it complete, then tailscale ip -4 will show the 100. address.
YOUOn your Mac: install Tailscale from tailscale.com/download (or the Mac App Store), open it, and sign in with the same account you just used (the "log in as [email]" it shows). Both devices on one account = one private network.
CHECKBack in the terminal window, paste tailscale ip -4. It prints your droplet's private address, starts with 100., write it down; the Mac app uses it next.

The Mac app

≈ 3 min

Now the part you originally asked about. The desktop app is a remote control: the agent keeps living on the droplet; this just gives you a nice window into it.

YOUDownload Hermes Desktop from hermes-agent.nousresearch.com and drag it into Applications, like any Mac app.
On first open, two prompts: (1) macOS asks to allow Local Network access, choose Allow (the app needs it to reach your droplet over Tailscale; the scary wording is macOS boilerplate). (2) If it offers to set up a model provider or a local Hermes, skip it, you're not building a new Hermes on your Mac, you're connecting to the one on your droplet (its brain is already set up). Look for a Skip or "connect to a remote/existing server" option, then do the next step.
Open it → Settings → Gateway → Remote gateway. The Remote URL is http:// + your droplet's Tailscale IP + :9119, for example http://100.114.22.9:9119 (use your number). Find the IP by clicking the Tailscale icon in your Mac's menu bar → hermes → the address starting with 100. (same one tailscale ip -4 prints on the droplet). Note: http, not https; keep the :9119. If hermes isn't in that list, finish tailscale up on the droplet first.
YOUA login form appears: username admin, password = your "Hermes dashboard" password from Part 5. Save / reconnect.
CHECKThe app shows your droplet's Hermes, its chat, sessions, and memory. Quit the app: nothing stops, because Hermes lives on the droplet. That's the whole idea.
Error "No inference provider configured" (or "Run hermes model")? The app is not connected to your droplet, it's using an empty local copy on your Mac. Don't run hermes model on the Mac. Instead: (1) on the droplet, confirm systemctl status hermes shows active (running) (that's Part 5, the hermes serve service must be running); (2) re-check Settings → Gateway → Remote gateway has the right http://100.x:9119 + admin/password and shows connected. Once connected, the droplet's Grok brain answers and the error's gone. If both of those are already fine, the error is coming from the droplet itself, its brain isn't set. Run hermes model on the droplet (never the Mac) to (re)configure the model provider.

Signal from your phone

≈ 8 min · optional

Text your agent from anywhere over Signal, end-to-end encrypted, free, and nothing has to run on your Mac. Signal links to your phone as a second device (like WhatsApp Web), so your phone stays in charge and you chat with the agent through Signal's own Note to Self, which is just a normal Signal chat window (your own self-chat) where the agent's replies appear.

Do I even need Signal? (the Hermes app already chats)

No, this part is optional. The Hermes app you set up in Part 7 already lets you chat with the agent, and it's the most private option: the connection runs over Tailscale (encrypted), the data stays on your own droplet, and no outside company is involved at all.

Signal only adds two things: chatting from your phone when you're away from your Mac, and push notifications so the agent can ping you (alerts, "task done"). If you're mostly at your Mac, skip this and just use the app. (You can even reach the dashboard from your phone privately, put Tailscale on your iPhone and open the URL in Safari, you'd just miss push notifications.)

Why Signal, over Telegram and iMessage (privacy)

Signal, your choice: End-to-end encrypted; the keys live on your own droplet and no outside company sits in the middle. Free, no Mac, 24/7. Trade-off: it links as a second device on your own number, so you chat via Note to Self and the agent can see your Signal. Want a fully separate assistant identity that's still end-to-end? Register Signal to a spare number instead (extra step + a dedicated phone number). If you don't otherwise use Signal, the agent is your only chat, that trade-off doesn't matter, so Note to Self is the easy call.
Hermex / Hermes Mobile (iPhone app), most private of all: A native iOS app that talks straight to your droplet over Tailscale, no messaging company anywhere, your data stays on your own hardware. For proactive alerts, pair it with ntfy (free, self-hostable push, no signup). Caveats: these are community apps (connect to your own server, but vet them, prefer open-source), you install Tailscale on your iPhone, and they connect to the hermes-webui component on the droplet.
Setting up Hermex: the app hands you a setup prompt. Paste it as a message to your agent in the Mac app (the agent chat), not into the raw droplet terminal; it's written as instructions for the AI, not shell commands, so bash would just error. The agent reads it, installs hermes-webui (port 8787), wires it to Tailscale, and replies with the URL + password to enter in Hermex. Verify auto-start: Hermes may block the agent from writing the systemd service file (a protected system path), so ask it to confirm systemctl is-enabled hermes-webui returns enabled, creating the service via its terminal tool if needed, so the WebUI survives reboots. Tailscale is already on the droplet (Part 6), so that step is a no-op. Safety: add a line telling it to keep 8787 Tailscale-only, do NOT add a ufw rule for 8787 (the tailscale interface is already allowed; opening 8787 would expose it on the public IP).
Telegram: Easiest to set up (a bot token, ~2 min) and a clean separate identity you DM like a contact, but bot chats are not end-to-end encrypted, so Telegram the company can read them. The same "a third party can read your messages" issue you avoided with Photon.
BlueBubbles (iMessage): Native iMessage through your own Mac and Apple ID, also end-to-end with no third party, and free. The catch: that Mac has to stay switched on to relay.
iMessage via Photon: Free tier and no Mac, but it's a hosted relay. Its privacy policy offers only in-transit and at-rest encryption (Photon holds the keys, not end-to-end), with no promise that your messages aren't read, retained, or used for training. Assume Photon can see them.

Prefer Hermex (iPhone app) instead of Signal?, the most private option

Hermex is a native iPhone app that talks straight to your droplet over Tailscale, no messaging company at all, everything stays on your own hardware. Rough steps:

1. Install Tailscale on your iPhone and sign in with the same account (so your phone joins the private network).

2. Install Hermex from the App Store. It shows a setup prompt, paste it into the Hermes agent on the droplet (run hermes on the console, that opens a chat screen; type/paste your message there, not at the plain root@hermes:~# prompt), not the Mac app. Two reasons: the WebUI has to install on the droplet, and the Mac app may run the task on a local copy instead. The droplet agent installs the hermes-webui component for you and keeps it behind a password + Tailscale-only.

3. Save the WebUI password the agent generates, then enter the server URL + password it gives you into Hermex.

For proactive push alerts (agent pinging you), add ntfy (free, self-hostable, no signup). With Hermex you don't need Signal at all. Caveat: Hermex is a community app, it connects only to your own server, but vet it (an open-source alternative is Hermes Mobile).

Signal setup steps (open only if you chose Signal)

YOUType your Signal phone number (the one your Signal app is registered to) in international format, the blocks below fill themselves in.
Signal number:Format: a plus sign, then country code, then the number, with no spaces or dashes. Stays on this page only.

Earlier, Hermes's first setup asked you to "select platforms" and you ticked Signal, that just flags it on. These steps are what actually make it work.

PASTEInstall Signal's engine on the droplet:

BLOCK J Java + signal-cli + QR tool

# Java + QR tool
apt-get -y install openjdk-17-jre-headless qrencode
# signal-cli (latest, with a fallback version if the lookup fails)
VER=$(curl -fsSL https://api.github.com/repos/AsamK/signal-cli/releases/latest | grep -oP '"tag_name":\s*"v\K[^"]+')
VER=${VER:-0.13.14}
curl -fsSL "https://github.com/AsamK/signal-cli/releases/download/v${VER}/signal-cli-${VER}.tar.gz" | tar xz -C /opt
ln -sf /opt/signal-cli-${VER}/bin/signal-cli /usr/local/bin/signal-cli
signal-cli --version

PASTEShow a QR code, then link it with your phone:
BLOCK K link Signal (scan the QR)
```
# start linking in the background, then draw a QR to scan
signal-cli link -n "HermesAgent" > /tmp/signal-link.txt 2>&1 &
sleep 6
LINK=$(grep -o 'sgnl://[^ ]*' /tmp/signal-link.txt | head -1)
[ -n "$LINK" ] && qrencode -t UTF8 "$LINK" || cat /tmp/signal-link.txt
```
On your phone: open Signal → Settings → Linked Devices → Link New Device, then scan the QR in the console. If the QR looks too big or cut off, zoom the browser out (⌘ −) and run Block K again.

Expect: your phone shows "HermesAgent" under Linked Devices, and the console returns to root@hermes:~#.

PASTEKeep Signal running 24/7 and let only you message the agent:

BLOCK L Signal service + connect Hermes

# run signal-cli as a 24/7 service
cat > /etc/systemd/system/signal-cli.service <<'EOF'
[Unit]
Description=signal-cli daemon
After=network-online.target

[Service]
ExecStart=/usr/local/bin/signal-cli --account +1XXXXXXXXXX daemon --http 127.0.0.1:8080
Restart=always
User=root

[Install]
WantedBy=multi-user.target
EOF
systemctl daemon-reload && systemctl enable --now signal-cli
# tell Hermes how to reach Signal, and that only your number may talk to it
cat >> ~/.hermes/.env <<'EOF'
SIGNAL_HTTP_URL=http://127.0.0.1:8080
SIGNAL_ACCOUNT=+1XXXXXXXXXX
SIGNAL_ALLOWED_USERS=+1XXXXXXXXXX
EOF
systemctl restart hermes

CHECKOn your phone, open Signal → Note to Self → send "are you there?". Hermes replies within a few seconds. Nothing? Paste systemctl status signal-cli hermes and check both say active (running) (press q to type again).

Save-point & first task

≈ 2 min

Freeze this fully-signed-in state so it can always be restored, then give your new employee its first assignment.

Take a snapshotiA snapshot is a frozen copy of the whole droplet, all software, settings, and sign-ins, that you can restore from if anything ever breaks., a one-click "save point" of everything you just set up. This step is optional and costs a little (~cents to ~$1/mo to store); the free alternative is the Cloudflare backup in Part 10. The difference: a snapshot restores the whole machine in one click, while the free backup restores your data after you rebuild the droplet. Do this only if you want the instant-restore convenience:
1. Go to cloud.digitalocean.com → Droplets → click hermes.
2. In the droplet's sidebar, click Snapshots.
3. In the "Take a snapshot" box, type hermes-fresh and click Take Snapshot (may say "Take Live Snapshot", fine).
4. Wait a few minutes until it appears in the list.
Storage is a few cents to ~$1/month. Restoring brings back everything, including your sign-ins.
In the Hermes chat (app or Signal Note to Self), send: "Introduce yourself, list the tools you can use, and confirm you can reach Claude Code and Grok Build."
CHECKIt answers, and names both coding agents. Done, you have a 24/7 AI employee. Now give it a role and a first real task, and let it work while you go do something else.

Free nightly backups to Cloudflare

≈ 5 min · $0

Every night the droplet sends an encrypted copy of everything that matters, Hermes's memory, your sign-ins, the Hermex WebUI, service files, and system config (/root, /opt, /etc), to storage in your own Cloudflare account (R2). Its free tier covers this easily, so neither the DO backup fee nor the Part 9 snapshot is required. (To restore: rebuild the droplet from this guide, then pull your data back from R2 with a single restic command.)

How this part flows: make the bucket + token on the Cloudflare site (steps below), type your keys into the 4 boxes, then paste Block H and Block I into the droplet console (in tmux), H connects the storage and creates the vault, I turns on the nightly schedule. These two blocks are terminal pastes, not for the Cloudflare site or the agent chat.

No Cloudflare account? Backup options without it

Cloudflare R2 is just the storage this guide happens to use, because its free tier makes nightly off-site backups free. You don't need Cloudflare specifically. Two alternatives:

Easiest (paid): DigitalOcean's own backups. In Part 1 you left "Enable automated backup plans" unticked. Tick it instead (about $5/month) and DigitalOcean snapshots the whole machine on a schedule with zero setup. You can then skip this entire part.

Free or cheap, off-site: any S3 storage. The same restic setup in Blocks H and I works with Backblaze B2 (10 GB free, then very cheap) or AWS S3. Just swap the endpoint, access key, and secret in Block H for that provider's values, and leave everything else the same.

YOUIn your Cloudflare dashboard: R2 Object Storage → Create bucket → name it hermes-backups. For Default storage class, choose Standard (not Infrequent Access) → Create.
Why Standard, not Infrequent Access: the free 10 GB tier only applies to Standard, IA is billed from the first byte. IA also adds a 30-day minimum charge and per-read retrieval fees, which our nightly backup would trip constantly (restic prunes and re-reads data often). IA only pays off for large, write-once, rarely-touched archives, the opposite of this.
YOUStill in R2: Manage R2 API Tokens → Create API Token → permission Object Read & Write → under "Specify bucket", choose only hermes-backups. Set TTL to Forever (this runs unattended, an expiring token would silently kill your backups). For Client IP filtering (optional): first run curl -4 ifconfig.me on the droplet to get its real outbound IP, and enter that, a leaked token then only works from your droplet. (Cloudflare appends /32 = that one exact IP, which is correct.) Leaving it blank is also safe, the token can only touch the hermes-backups bucket either way, so a leak can't reach anything else in your Cloudflare. If unsure, leave it blank to avoid an "Access Denied" from an IP mismatch. Then Create. Keep this page open, it shows an Access Key ID, a Secret Access Key, and an endpoint URL containing your Account ID (the long code before .r2.cloudflarestorage.com). ⚠️ Use the Access Key ID (a hex string), NOT the "Token value" that starts with cfat_; that's a general Cloudflare token restic doesn't use, and grabbing it by mistake causes a 400 Bad Request.
YOUFill these four boxes, Block H below customizes itself. The backup password is one you invent (make it a new, distinct password, not reused from your dashboard or anything typed into a chat).
⚠️ This backup password is the encryption key for your backups, save it and never lose it. restic warns: "Losing your password means that your data is irrecoverably lost." Here "data" = everything backed up to R2, Hermes's memory, your sign-in tokens, all configs, and the Hermex WebUI. Without this password, those backups can't be decrypted by anyone, including you. Save it in your password manager (Apple Keychain is fine) as "Hermes backup password".

Access Key ID:

Secret Access Key:

Account ID:

Backup password (invent one):All four stay on this page only, they're just filled into the blocks for you.
PASTEInstall resticirestic is a free, open-source backup program. It encrypts everything before it leaves your droplet and only uploads what changed, so backups stay private and small., then connect it to your storage:
BLOCK H connect the droplet to your Cloudflare storage
```
apt-get -y install restic
cat > /root/.restic-env <<'EOF'
export AWS_ACCESS_KEY_ID=PASTE-ACCESS-KEY-ID
export AWS_SECRET_ACCESS_KEY=PASTE-SECRET-ACCESS-KEY
export AWS_DEFAULT_REGION=auto
export RESTIC_REPOSITORY=s3:https://PASTE-ACCOUNT-ID.r2.cloudflarestorage.com/hermes-backups
export RESTIC_PASSWORD=PASTE-BACKUP-PASSWORD
EOF
chmod 600 /root/.restic-env
. /root/.restic-env && restic init
```
Expect: "created restic repository … at s3:https://…", the vault exists.

Got "Access Denied"? Troubleshoot here
Most often the token's Client IP filtering doesn't match the droplet's real outbound IP (the reserved IP is not the egress IP). Run curl -4 ifconfig.me on the droplet, then edit the token to that IP, or just remove the IP restriction, and re-run . /root/.restic-env && restic init. If there was no IP filter, it's the keys: run cat /root/.restic-env and confirm AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY are real values (not PASTE-… placeholders from pasting Block H before filling the boxes) and match your current token. Recreating a token changes its keys, so if you edited/recreated the token, refill the 4 boxes and re-paste Block H so the file matches. Also open the token itself and confirm: Permissions = Object Read & Write (not any "Read only", restic must write); Client IP filtering is empty; and the bucket name matches in all three places, your real bucket, the token's bucket scope, and Block H's RESTIC_REPOSITORY path, spelled identically (a stray hermes-backup vs hermes-backups = Access Denied). A wrong permission level or a leftover IP filter also show up as "Access Denied." (To realign restic to your actual bucket name: sed -i 's#hermes-backups#YOUR-BUCKET#g' /root/.restic-env then re-init.)

Error mentioning PASTE-ACCOUNT-ID (or another PASTE-…) / a TLS handshake failure? That means a box above was still empty when you copied, the block has placeholder text, not your keys. Fill all four boxes, re-copy this block, and paste again (safe to re-run, it overwrites the file).

PASTE

BLOCK I run it automatically every night

# keep backups lean, skip regenerable dirs (node_modules, caches, virtualenvs)
cat > /root/.restic-excludes <<'EOF'
**/node_modules
**/.cache
**/.npm
**/__pycache__
**/.venv
**/venv
EOF
cat > /etc/systemd/system/hermes-backup.service <<'EOF'
[Unit]
Description=Nightly backup to Cloudflare R2

[Service]
Type=oneshot
ExecStart=/bin/bash -c '. /root/.restic-env && restic backup /root /opt /etc --exclude-file /root/.restic-excludes --exclude-larger-than 200M && restic forget --keep-daily 14 --keep-weekly 8 --prune'
EOF
cat > /etc/systemd/system/hermes-backup.timer <<'EOF'
[Unit]
Description=Nightly backup timer

[Timer]
OnCalendar=*-*-* 09:00:00
Persistent=true

[Install]
WantedBy=timers.target
EOF
systemctl daemon-reload && systemctl enable --now hermes-backup.timer
systemctl start hermes-backup

CHECKPaste . /root/.restic-env && restic snapshots, a table with one row appears: your first backup is in Cloudflare. From tonight it repeats itself at ~2am Pacific, keeping two weeks of nightlies. If we ever need it: rebuild from this guide + one restore command with me, and memory and sign-ins come back.

That's the whole build. You now have a private, always-on AI agent living on a computer you own, reachable from your Mac and your phone from anywhere, running on the AI provider you chose, and quietly backing itself up every night. Why this makes you far more productive: it works while you sleep, runs multi-step jobs on its own and several at once, remembers your context so you never re-explain, and pings your phone when it needs you. You stop doing every task yourself and start delegating, which is the whole point.

From here the real payoff begins. Give it a role, a schedule, and a first real task, and let it work while you go do something else. That is the jump from using AI to having AI work for you.

Build Your Own AI Agent: A Private, Always-On Assistant You Actually Own

What makes this powerful

Read this first

Create the cloud computer

Open its screen & prepare it

Install the three programs

Sign in to Claude & Grok

Turn Hermes on, permanently

Private line to your Mac

The Mac app

Signal from your phone

Save-point & first task

Free nightly backups to Cloudflare