← Back to blog list

The Personal Agent Era: Why a VPS Is the Best Companion?

From cloud functions to laptops to local Docker — why a VPS with a public IP is the best vehicle for personal AI agents. One year of hands-on lessons.

·7 min read·by TOMMi

"In the Agent era, everyone needs a tiny machine that runs 24/7." — a netizen, early 2026

For the past year I've been running my entire tommiai.xyz site, the AI customer support, and the crypto trading bots on a single Tokyo VPS — 4GB RAM, ~$5/month, no reboot in 200+ days. Here's why I think a VPS is the strongest companion of the personal Agent era.

1. The Personal Agent Era Needs a "Permanent Home"

Let me set the scene — what's the real difference between today's AI Agent and a traditional chatbot?

Traditional Chatbot

  • Gone after you close the tab
  • No memory
  • No "body" to touch the outside world

Personal Agent

  • Online 24/7 — sends your email at 3am
  • Has memory — remembers you across sessions
  • Has hands and feet — can touch GitHub, email, databases, the browser
  • Can orchestrate — watches markets, posts tweets, replies to email, runs batch jobs

An AI that's always on, remembers everything, and can act on the world needs a "home" — a server where its long-running process lives.

2. Why VPS Beats Every Other Option

Here are all the options I tried this year, side by side:

Option Comparison

| Option | Monthly Cost | 24/7 Online | Stable | Flexible | Private | Overall | |---|---|---|---|---|---|---| | VPS (Vultr / AWS Lightsail) | $5–7 | ✅ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | | Cloud Functions (Cloudflare Workers) | Free | ✅ | ⭐⭐⭐⭐ | ⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | | Laptop always-on | $0 | ⚠️ Sleep breaks it | ⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐ | | Home Docker (Raspberry Pi) | ~$1/power | ✅ | ⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | | Third-party SaaS (Coze / Manus) | $0–15 | ✅ | ⭐⭐⭐⭐⭐ | ⭐ | ❌ Closed source | ⭐⭐ |

In Detail

VPS: ⭐⭐⭐⭐⭐ Strongly Recommended

Pros:

  • Real 24/7 — no sleep, no hibernate, no shutdown
  • Public IP — Agent can call any external API; you can reach it from anywhere
  • Stable — datacenter network, UPS, redundant power
  • Full control — you're root, install whatever
  • Observable — your own logs, your own monitoring
  • Cheap — 4GB starting at $5/month, less than a coffee

Cons:

  • ⚠️ Need to know a bit of Linux (but really only 10 commands)
  • ⚠️ Security is on you (fail2ban + SSH key is the baseline)
  • ⚠️ Backups are on you (3-2-1 rule)

Cloud Functions (Cloudflare Workers / Vercel)

Pros: Generous free tier, auto-scaling

Cons:

  • No persistent storage — every cold start is amnesia
  • ❌ Hard to install large dependencies (AI models don't fit)
  • ❌ Execution time caps (Workers 30s, Vercel 10s)
  • ❌ Want SQLite? Hard. Want Docker? No.

Agents need state, memory, long-running tasks — cloud functions are the stage for "stateless little tools", not a home for Agents.

Laptop Always-On

Pros: Zero cost, full control

Cons:

  • Sleep/hibernate breaks it — closing the lid = offline
  • ❌ Network down = offline
  • ❌ Power outage = offline
  • ❌ A laptop just isn't designed to run 24/7

Raspberry Pi / Home Server

Pros: Full control + 24/7 + data at home

Cons:

  • No public IP — sitting behind carrier-grade NAT
  • Need tunnel (Cloudflare Tunnel / FRP) — added complexity
  • Home network is unstable — your parents watching TV adds jitter
  • Power outage = permanently offline — no UPS

A home server is great for "local compute" workloads, not for "Agent serving the outside world".

Third-Party SaaS (Coze / Manus / Devin)

Pros: No ops, works out of the box

Cons:

  • Data isn't yours — trade secrets going to someone else's cloud
  • Capability limited — can't install custom skills
  • Can get banned — policy shifts and the service goes away
  • Expensive long-term — once you're hooked, $75+/month

3. What's the "Best Practice" for VPS?

If you're going VPS, here's my minimum viable stack:

Sizing

| Dimension | Recommendation | |---|---| | Provider | Vultr / AWS Lightsail / DigitalOcean / Aliyun Lightweight | | Datacenter | Tokyo (Asia) / Singapore / LA (Americas) | | Specs | 4GB RAM / 80GB SSD / 4TB transfer ($5–7/month) | | OS | Ubuntu 24.04 LTS (beginner-friendly) |

Must-Have Software

# 1. SSH security
sudo apt install fail2ban
# Change SSH port + disable password login + use SSH key

# 2. Firewall
sudo ufw allow 22,80,443
sudo ufw enable

# 3. Docker (run Agent in containers)
curl -fsSL https://get.docker.com | sh

# 4. Monitoring
# 5-min cron + Telegram alerting

What to Run

# Your Agent runs inside Docker
docker run -d \
  --name hermes \
  --restart always \
  -p 8642:8642 \
  -v /home/tommi/.hermes:/root/.hermes \
  hermes-agent:latest

Backup

3-2-1 rule:

  • 3 copies
  • 2 different media
  • 1 offsite
# Weekly cron to object storage (R2 / S3)
rclone sync /home/tommi/.hermes r2:tommi-backup

4. My Real Architecture

Behind tommiai.xyz, four core services all run on a single 4GB VPS (~$5/month):

┌─────────────────────────────────────────┐
│  VPS: VM-TYO1 (Tokyo, 4GB RAM)         │
│                                          │
│  ┌────────────────────────────────────┐ │
│  │  Hermes Gateway (PID 826815)       │ │
│  │  - 200+ days uptime                │ │
│  │  - Multi-model routing             │ │
│  │    (MiniMax / Opus)                │ │
│  │  - 115+ Skills                     │ │
│  └────────────────────────────────────┘ │
│                                          │
│  ┌────────────────────────────────────┐ │
│  │  Cloudflare Tunnel (PID 840455)    │ │
│  │  - 4 QUIC edge nodes               │ │
│  │  - 5-min watchdog cron             │ │
│  └────────────────────────────────────┘ │
│                                          │
│  ┌────────────────────────────────────┐ │
│  │  Freqtrade × 3 Bots                │ │
│  │  - BTC/ETH/SOL/HYPE/BNB            │ │
│  │  - Live trading for 6 months       │ │
│  └────────────────────────────────────┘ │
│                                          │
│  ┌────────────────────────────────────┐ │
│  │  Nginx + acme.sh                   │ │
│  │  - Auto-renew HTTPS                │ │
│  │  - Reverse proxy                   │ │
│  └────────────────────────────────────┘ │
└─────────────────────────────────────────┘

Monthly total: $5 (VPS) + $0 (Cloudflare free tier) + LLM API costs on top

5. Which VPS — 2026 Real-World Picks

| Provider | Pros | Cons | Rating | |---|---|---|---| | Vultr | 25+ global DCs, hourly billing, destroy/rebuild anytime | Console feels a bit slow | ⭐⭐⭐⭐⭐ | | AWS Lightsail | Stable, great docs, full ecosystem | Slightly pricier than Vultr | ⭐⭐⭐⭐ | | DigitalOcean | Veteran, clear docs, student-friendly | Fewer DCs | ⭐⭐⭐⭐ | | Aliyun Lightweight | Fast in China, ICP-friendly | Real-name required, overseas needs workaround | ⭐⭐⭐ (in China) | | BandwagonHost | Cheap ($50/year) | Average performance | ⭐⭐⭐ |

I picked Vultr Tokyo: physically close to mainland China + good international bandwidth + $5/month + 5-minute provisioning

6. The Future of VPS — Edge VPS?

VPS is evolving. The 2026 trends:

1. Edge VPS

Cloudflare Workers now has Workers + Durable Objects = edge stateful storage. Combined with D1 / R2 / KV, you can build "globally distributed Agent clusters".

Personal Agents running on the edge node closest to the user, latency < 50ms.

2. AI-Optimized VPS

Some providers now sell GPU VPS (RunPod / Vast.ai / Lambda Labs) — rent an A100/H100 by the hour, run local LLMs.

Future: $1.5/hour for an A100, run a 70B model, fully local inference, data never leaves the VPS.

3. Self-Hosted LLM VPS

Providers ship VPS with "Ollama + Open WebUI pre-installed" (like EasyPanel / OpenHands) — beginner-friendly.

One-click private ChatGPT, data stays in your hands.

7. Conclusion

In the personal Agent era, everyone needs a VPS.

  • Cloud functions are too stateless to host an Agent
  • Laptops are unstable
  • Raspberry Pis have no public IP
  • SaaS has bad privacy

VPS hits the sweet spot:

  • ✅ 24/7 online
  • ✅ Public IP
  • ✅ Full control
  • ✅ Data private
  • ✅ $5–7/month, cheaper than a coffee

My 2026 advice to indie developers: stop sinking time into cloud functions, get a $5/month VPS running — and you'll discover "oh wow, an Agent can really work for me 24/7".

8. Getting Started


What do you use to run your Agent? Chat on Telegram, or drop a comment below.

Related reading: