Giving Ultron Super Powers: VPS, Hermes, and N8N

Or: How I Turned a Local AI Into a Full-Time Butler, Nanny, and Accountant

Call me crazy, but a local AI sitting on your desk is cool. A local AI that can act on your behalf is dangerous — in the best possible way.

As a humble human who built Ultron from scratch — remember the moment I realized Apple Silicon’s unified memory was the missing piece? The one that finally made DeepSeek run fast, silent, and cool on my Mac? — let me walk you through how I connected everything and turned a language model into a real-world assistant.

Putting Ultron on the Internet

A model running on localhost is useful for experiments. A model running behind a proper VPS with nginx and Tailscale is useful for real life.

Here’s the split:

The brain (my Mac): Apple Silicon’s unified memory is where the model lives — fast, silent, and cool. It serves the REST API and handles embeddings.
The front door (a VPS): A headless Ubuntu server — cheap, reliable, always on — with nginx and SSL.
Tailscale: The private link that connects them. The VPS reaches the model over a 100.x private IP, never the public internet.

The setup is dead simple. nginx sits in front of everything and proxies to api.vedabyte.ai with Let’s Encrypt certs. The model lives on the Mac, reachable only over Tailscale. No cloud dependency, no API keys, no rate limits.

Giving Ultron Agency with Hermes

A model that just generates text is a fancy autocomplete. A model that can read files, search the web, run code, and send messages is an agent.

That’s where Hermes Agent comes in — the AI agent framework that runs Ultron with actual tool access. Hermes gives Ultron:

File system access — read, write, search files
Web access — search and extract content
Terminal access — run commands and scripts
Discord gateway — send notifications, respond to queries
Email integration — read and send via ProtonMail Bridge

Suddenly Ultron isn’t just a chat bot. It’s a sysadmin, a researcher, a content creator, and a notification system all in one.

The Automation Brain: N8N

N8N is the workflow automation engine that ties everything together. It runs in Docker on the same VPS, exposed through the same nginx proxy — and it’s where Ultron truly earns its keep. N8N is a visual canvas that connects any tool to any other, with the model doing the thinking in between.

Some workflows that are genuinely worth building:

Automated developer — spin up staging environments, run test suites, and deploy on a webhook or a timer instead of babysitting the terminal.
Automatic PR reviewer — watch a repository, and have Ultron read pull requests and drop suggestions right in the diff.
Daily news digest — gather AI and tech headlines, summarize them with the model, and drop a clean brief in Discord or your inbox every morning.
Incident briefings — collate overnight alerts and warnings into a summary you read with your coffee.
Content scheduler — draft posts, queue them, and publish on a schedule.

The point isn’t any single workflow. N8N hands you the tools; what you build is up to you.

Your Interface, Your Choice

The Hermes dashboard is what the script installs by default. Fair warning: it has a retro, techie vibe — built for people who like poking at terminals, not everyone. If that aesthetic speaks to you, wonderful.

And here’s what changes everything: you don’t even need a dashboard. If none of those appeal to you, just tell Ultron where you actually want to talk — Discord, Telegram, WhatsApp, whatever you already live in every day — and it’ll walk you through wiring it up. It’s like having your own IT guy who not only builds the plumbing but teaches you to use it.

Making It a Day-to-Day Assistant

This is what turned Ultron from a neat server into something genuinely useful around the house.

Using system prompts directly in the interface, I created dedicated AI personas:

Writing partner — helps draft emails, polish prose, and organize thoughts
Research assistant — searches the web and summarizes findings into plain language
Cookbook helper — suggests recipes from whatever’s in the fridge
Travel planner — maps out itineraries and compares options

No coding required. Just a prompt and a model. Everyone in the house can interact through the web interface on their devices. The model runs locally — no data leaves the house, no subscription fees, no reliance on public APIs.

The Final Stack

Here’s what Ultron looks like today:

   ┌────────────────────┐
   │   MAC (the brain)  │
   │  Model + DeepSeek  │
   │  (over Tailscale)  │
   └─────────┬──────────┘
             │ 100.x private network
             ▼
   ┌──────────────────────────────────┐
   │           VPS (front door)       │
   │                                  │
   │  nginx (SSL)                     │
   │   ├─ Hermes dashboard       ← techie │
   │   ├─ Hermes agent → Discord          │
   │   ├─ Telegram / WhatsApp             │
   │   ├─ n8n (automation)                │
   │   └─ ProtonMail (email)              │
   └──────────────────────────────────┘

The front door runs on a headless Ubuntu VPS behind a single nginx reverse proxy. The brain runs on the Mac, connected over Tailscale. Whether you talk to it through a web dashboard, a chat app, or just ask it to set one up — everything talks to each other.

The Journey

This series started with a 1 AM shopping spree fueled by token anxiety. It went through hardware mistakes, Docker facepalms, Apple revelations, and tooling discoveries. And it ended with a fully connected AI system that sorts my email, drafts my messages, researches topics, posts news to Discord, and tracks my budget.

From token anxiety to a full AI stack — that’s the adventure of creating Ultron.

And the best part? Anyone can do it.

➡️ Build your own Ultron → — the full setup scripts and step-by-step guide to build your own Ultron, from scratch.

And because I’ve decided to share them with you — here they are, courtesy of LloydsByte LLC.