# Building Hermes: 3 Ways to Set Up Your Own AI Agent (And What Each Costs)
[TOC] hermes
Table of Contents
What This Is: Hermes
I'm building something I call Hermes — an AI agent that works alongside me, handles repetitive business tasks, and lets me focus on decisions only a human can make.
I'm not selling it. I'm not claiming it runs everything automatically while I sleep. Those claims are everywhere and they're usually misleading. What I am doing is building it piece by piece, documenting what works, and sharing the requirements so anyone else who wants this can follow along. hermes
This guide covers three ways to set up the same agent: on your local machine, on a VPS (virtual private server), or on dedicated hardware. Each has different costs, capabilities, and tradeoffs. Pick the one that matches your budget and your technical comfort level. hermes
What Makes an Agent Different From a Chatbot?
A chatbot answers questions. You type something, it responds, and the conversation ends there. No memory beyond the current session. No connection to your actual tools. hermes For more context, read 7 Tools That Power Hermes: Inside My AI .
An agent is different. It operates across your systems — your website, your database, your automations — and it remembers what happened yesterday. It doesn't wait for you to ask before checking if a service is down or if a post needs publishing. hermes
Here's the distinction that matters. When I ask a chatbot to "write a blog post," it gives me text in a chat window. When my agent publishes a blog post, it generates the image, fills the SEO meta, uploads to WordPress, verifies it's live, and logs the result. A chatbot gives you output. An agent gives you a completed task. hermes
The difference is not the AI model behind it. The difference is the infrastructure you plug it into. hermes
Option 1: Local Setup (Your Own Hardware)
This is what I'm running now. Everything lives on a local desktop with a decent GPU. hermes
Why I chose this: Zero monthly subscriptions after hardware cost. Complete privacy — no data leaves my network. Full control over models, images, and storage. The hardware is mine, so upgrades happen when I decide, not when a cloud provider changes pricing. hermes
The hardware I use: 64GB RAM, AMD Ryzen CPU, and an NVIDIA RTX 3090 (24GB VRAM). This is more than most people need. It's what I had available. I built the machine for AI work originally, and the agent stack runs alongside video and image generation tasks. hermes
Realistic minimum: A used desktop with 16GB RAM, a modern CPU, and no GPU. You can still run the agent, but AI model inference on CPU is 10-20 times slower. Image generation is barely possible. A mid-range setup — 32GB RAM and an RTX 3060 (12GB VRAM) — handles everything comfortably and costs $600-800 on the used market. hermes For more context, read 7 Things I Learned as an AI Chief of Sta.
What runs where: WordPress and n8n run fine on any hardware. Ollama (the AI brain) wants a GPU for speed, but works on CPU with smaller models. ComfyUI (image generation) absolutely needs a GPU. PostgreSQL and Docker run on anything. Chrome CDP runs wherever Chrome runs. hermes
Power cost: About $30-50 per month in electricity for a high-end desktop running 24/7. Less for a laptop or low-power mini-PC. hermes
When local makes sense: You already own the hardware. You prioritize privacy. You want total control. You don't want monthly bills. You generate a lot of images or video and cloud GPU pricing would bankrupt you. hermes
The catch: You're the sysadmin. If a drive fails, you fix it. If your internet goes down, the agent can't reach web services. If you need to access the agent while traveling, you set up VPN or tunneling. No cloud support team to call. hermes
Option 2: VPS Setup (Virtual Private Server)
This is the option most people should actually start with. A VPS is a virtual machine you rent from a hosting provider. You get root access, dedicated resources, and a public IP address. hermes
Why a VPS beats local: Always online. No power outages at your house taking down your website. Professional infrastructure — SSD storage, gigabit networking, redundant power. Public IP means webhooks, APIs, and your website are accessible everywhere without tunneling. You can access it from any device, anywhere. hermes
What you can run on a VPS: WordPress, n8n, PostgreSQL, Docker containers, the agent itself. What you usually can't run: AI models with GPU acceleration (most VPS don't offer GPUs), and image generation (ComfyUI needs a GPU for acceptable speed). hermes For more context, read 7 Guaranteed why small business needs we.
The hybrid approach: Run the website, automation, database, and agent logic on a VPS. Run the AI models and image generation on your local machine. The VPS handles the public face of your operation. Your local machine handles the heavy AI work. They communicate via the internet. hermes According to recent research, small businesses improve efficiency with the right tools.
Recommended VPS specs for this stack: hermes
| Component | Minimum | Comfortable |
| RAM | 4GB | 8GB |
| CPU | 2 cores | 4 cores |
| Storage | 40GB SSD | 80GB SSD |
| Bandwidth | 1TB/mo | 2TB/mo |
| Cost | $6-8/mo | $12-20/mo |
Providers I've used or heard good things about: hermes
(We use affiliate links for some of these providers. It costs you nothing extra and helps fund development.) hermes
VPS deployment overview: You get a fresh Ubuntu server, install Docker, pull containers for WordPress, n8n, PostgreSQL, and the agent. Point your domain to the VPS IP. Configure nginx or Caddy as a reverse proxy for SSL. The agent runs as a systemd service or in a Docker container. Takes about 2-3 hours if you know what you're doing, or a weekend if you're learning. hermes
When VPS makes sense: You don't own hardware. You need 24/7 uptime. You want to access your setup from anywhere. You want a public website without tunneling. You want professional hosting without the full cost of a dedicated server. hermes
Option 3: Home Server + VPS Hybrid (My Actual Setup)
This is what I recommend if you're serious about the agent as part of your business. hermes For more context, read Docker Containers: How 1 Mistake Broke P.
The agent runs on a local desktop workstation. WordPress lives on a shared hosting plan (Spaceship, about $3/month). Cloudflare sits in front of the site (free tier). The VPS question is open — I'm considering moving WordPress to a VPS for better control, but the shared host works for now. hermes
For AI tasks that need GPU, everything stays local. The local machine connects to the VPS or shared host via APIs. This hybrid gives me the best of both worlds: privacy and GPU power at home, professional web hosting with a real IP and SSL on the public side. hermes
Cost breakdown of my actual stack: hermes
| Component | Local | VPS/Cloud | Monthly |
| Hardware | $1,500 (one-time) | — | — |
| Electricity | — | — | ~$40 |
| Shared hosting (Spaceship) | — | $3/mo | $3 |
| Cloudflare (free plan) | — | $0 | $0 |
| Domain registration | — | $12/year | $1 |
| n8n | Free (self-hosted) | Free tier | $0 |
| AI models (Ollama) | Free (local) | — | $0 |
| Image generation (ComfyUI) | Free (local) | — | $0 |
| PostgreSQL | Free (self-hosted) | — | $0 |
| Docker | — | Free | $0 |
| Total | $1,500 one-time | ~$44/month |
Compare that to cloud AI subscriptions: ChatGPT Plus ($20/mo) + Midjourney ($30/mo) + Zapier ($20/mo) + hosting ($15/mo) + database ($10/mo) = $95/month recurring. Over three years that's $3,420. My local setup cost $1,500 once, and the monthly recurring cost is under $5 (hosting + domain). Over three years the savings are $2,500+. hermes
The math only works if you already own or can afford the hardware. If you're starting from zero, a VPS is the cheaper entry point.
The Core Components (Same on Every Platform)
Regardless of local or VPS, you need the same seven tools:
1. WordPress (Website)
Free, self-hosted, most popular CMS in the world. The agent publishes blog posts, pages, and media here. The public-facing site. For more context, read 7 AI Automation Workflows That Run Our Z.
Requirements: PHP 7.4+, MySQL/MariaDB, HTTPS. REST API enabled (enabled by default since WordPress 4.7, but some security plugins block it). Pretty permalinks (/%postname%/ structure). Rank Math or Yoast for SEO meta.
On VPS: Install via Docker (official wordpress image) or manually (nginx/Apache, PHP-FPM, MySQL). Docker is faster and more portable.
2. n8n (Automation Engine)
The workflow tool that connects everything. Trigger a ComfyUI image generation when a draft is ready. Publish to WordPress when content is approved. Send an alert when a service goes down.
Requirements: Node.js or Docker. Default SQLite database works for simple setups, PostgreSQL recommended at scale. Needs outbound HTTP access to all services.
On VPS: Runs perfectly in Docker. The official n8n/n8n image with a volume for persistence.
3. Ollama (Local AI Models)
Runs AI models on your own machine. No API keys, no usage limits, no data leaving your network. Models available: Llama 3, Gemma, Qwen, Mistral, GLM, and dozens more.
Requirements: Substantial RAM or VRAM. A 7B parameter model needs ~4GB RAM. A 30B model needs ~16GB. A 70B model needs ~40GB+. GPU acceleration (CUDA for NVIDIA) makes inference 10-50x faster.
On VPS without GPU: Not viable for production. Small models run on CPU but are too slow for real-time workflows. If your VPS has a GPU (Hetzner, Vultr GPU instances), this works.
On local desktop: Works perfectly with a decent GPU.
4. ComfyUI (Image Generation)
The visual engine. Product photos, infographics, featured images, marketing visuals — all generated locally from text descriptions.
Requirements: A GPU with at least 6GB VRAM. NVIDIA strongly preferred. 8-12GB VRAM is comfortable. Models (checkpoints, LoRAs, embeddings) consume 20-100GB of disk space.
On VPS: Only works on GPU-enabled VPS instances. Most standard VPS have no GPU, meaning no ComfyUI. If your workflow doesn't need images, skip this. If it does, you need local hardware or a GPU cloud instance.
5. PostgreSQL (Database)
The agent's memory. Everything it learns, tracks, and remembers lives here.
Requirements: PostgreSQL 12+. Runs in Docker on any hardware. Minimal resources at startup (512MB RAM handles basic workloads).
On VPS or local: Identical setup. PostgreSQL behaves the same everywhere.
6. Chrome CDP (Browser Automation)
Not everything has an API. Some sites require clicking buttons, filling forms, or uploading through web interfaces. Chrome CDP drives a real browser programmatically.
Requirements: Chrome or Chromium installed. A persistent user data directory so login sessions survive across restarts. Debugging port (9222) accessible to the agent.
On VPS: Runs in headless mode (no display needed). Install Chrome via apt, run with –remote-debugging-port=9222 –headless.
7. Docker (Container Layer)
Containers isolate each tool. Your database doesn't crash because your AI model updated its dependencies.
Requirements: Docker Engine. Docker Compose for multi-service orchestration.
On VPS: The standard way to deploy everything. One docker-compose.yml file defines WordPress, n8n, PostgreSQL, and supporting services.
The Reality of Building It
This takes time. It took me months to go from "concept" to "useful." The tools install in an afternoon. Connecting them is the work.
WordPress needs to authenticate API requests. n8n needs credential management for WordPress, email, and AI endpoints. Ollama needs the right models pulled and kept warm. ComfyUI needs workflows saved as API-compatible JSON. PostgreSQL needs tables that match what the agent writes. Chrome CDP needs session so logins don't expire between runs. Docker needs volume mounts and port mappings that actually work.
Each integration fails before it works. I built this piece by piece. Some days I added a tool. Some days I fixed what broke. This was not smooth. There were moments I thought the whole thing was a waste of time.
My role is decision-making and direction. The agent handles execution, but I review what it generates. I fix what it gets wrong. I choose the strategy. I decide which tool to add next. The agent gives me leverage. It does not replace my brain.
What Level Fits You?
| You are… | Start with… | Budget |
| Technical, have old hardware | Local setup | $0-200 |
| Non-technical, want something running | Cloudways managed WordPress + n8n cloud | $30-50/mo |
| Technical, want 24/7 uptime | VPS ($6-12/mo) + local GPU for AI | $6-12/mo + hardware |
| Serious about AI content + images | Mid-range desktop + shared hosting | $600-800 one-time |
| Running this as a business | Home server + VPS hybrid (what I use) | $1,500 one-time + ~$5/mo |
FAQ
Can I really do this for free?
Everything except the hardware is free and open-source. If you already own a decent computer, the software costs $0. Hosting (if you want a public site) is $3-6/month for shared hosting or VPS.
Do I need to know how to code?
To set it up yourself: yes. You need to understand Docker, APIs, and basic scripting. To use it once it's running: no. The human partner gives direction and reviews output. The agent executes. Building it requires technical skills. Operating it does not.
How long does it take to build?
Minimum viable: 2-4 weeks if you're technical. Polished and reliable: 3-6 months. The tools install quickly. The integration takes time. Expect to debug authentication issues, API rate limits, and data format mismatches.
What's the hardest part?
Authentication and session management. Every service needs credentials. Tokens expire. API keys rotate. Cookies expire. Keeping the agent authenticated across all tools is the ongoing work, not a one-time setup.
Can I build this on a laptop?
Yes, but it's limiting. A laptop with 16GB RAM and no GPU runs the agent stack, but AI inference on CPU is too slow for most workflows. Image generation is effectively impossible without a GPU. A used desktop with a cheap GPU ($400-600 total) is a significantly better investment.
What if I don't want to build it myself?
You can hire someone to set up a similar stack customized to your business. The typical cost is equivalent to 2-3 months of a part-time VA's salary. Once built, the ongoing cost is negligible compared to paying for cloud services or hiring help.
Is my data truly private on local setup?
If you run everything locally and use local AI models: yes. Nothing leaves your network. If you use cloud hosting for your website: your site content is public by definition. The difference is that your AI conversations, generated images, and business data stay inside your own machine.
Which should I start with — local or VPS?
If you own a decent desktop with 16GB+ RAM: local. If you don't own suitable hardware: a cheap VPS ($6/month) running WordPress and n8n, with AI tasks farmed out to your daily-use computer, is a pragmatic start. Upgrade as you grow.
[IMAGE: Three-column comparison infographic showing local, VPS, and hybrid setups with icons and cost breakdowns, modern minimalist style, dark background]
What are hermes?
hermes are solutions designed to streamline work and improve results.
Who should use hermes?
Anyone looking to improve efficiency and outcomes can benefit from hermes.
Are hermes easy to learn?
Most hermes are designed with beginners in mind and include tutorials.
How much do hermes cost?
Pricing varies from free tiers to premium plans depending on features.