Overview
I had two HP Pavilion 15-au laptops, each with something wrong with it. Instead of binning them, I combined the good parts into one working machine, installed Linux Mint, stripped it back to run headless, and now use it as an always-on server I reach over SSH from my MacBook. I built it to learn Linux administration, Docker, and how deployment works. The full setup lives in my home-server repo.
It currently runs a hardware monitor and a local AI backend, with a scraper planned next.
Hardware
Two donor laptops, same chassis family so all parts swap:
- HP Pavilion 15-au077sa — older (6th-gen i5-6200U), failing battery, but a nicer screen/keyboard, an SSD, and 8 GB RAM.
- HP Pavilion 15-au181sa — newer (7th-gen i5-7200U), better battery, 8 GB RAM, but a slow hard drive.
The goal was one machine with the newer CPU, the better screen and keyboard, both RAM sticks, and the SSD. Final build:
- CPU: Intel Core i5-7200U (2 cores / 4 threads)
- RAM: 16 GB DDR4 (dual-channel)
- Storage: 256 GB SanDisk SSD
- OS: Linux Mint 22.3 Cinnamon
The Build
The CPU is soldered to the board, so getting the newer processor meant swapping the whole motherboard into the older laptop’s body. I moved across the good battery, both RAM sticks, and kept the existing SSD; the old board and slow drive went to the spare-parts pile. It booted first time.
I chose Linux Mint Cinnamon because it’s stable, beginner-friendly, and built on Ubuntu. Post-install I set up SSH for headless access, enabled the ufw firewall (deny incoming by default), lowered swappiness to 10, and added Timeshift snapshots and TLP power management.
Services
| Service | Purpose | Port | How it runs |
|---|---|---|---|
| Glances | Hardware monitoring (CPU, RAM, disk, temps) | 61208 | Docker |
| Ollama | Local AI model backend | 11434 | Docker |
| Open WebUI | Browser interface for Ollama | 8080 | Docker |
The machine is headless, so I can check its health from my laptop using Glances, a web dashboard of load, memory, disk, network, and temperatures. It runs as a Docker container set to restart automatically. Alternatively, I can SSH in and run btop.
Architecture
MacBook Air HP Pavilion (server)
---------- --------------------
Terminal / SSH --------> Linux Mint (headless)
Web browser --------> |
LAN +-- Docker
|
+-- Glances (61208)
+-- Ollama (11434)
+-- Open WebUI (8080)All work happens from the Mac: admin over SSH, services viewed in the browser over the LAN.
Problems & Solutions
The “good” battery was also worn out. After the build it sat at 100% then dropped fast. upower showed why: full capacity had fallen to ~14 Wh against a ~41 Wh design figure, and the controller had quietly rewritten its own baseline to report “100% health.” Both batteries were near end of life.
Glances ran but served nothing. The container showed as running but nothing loaded. curl to localhost on the server showed the port wasn’t responding, so the problem was the container, not the network. The web-server option was being passed as an environment variable the image ignored; recreating the container with the web-server flag as a command fixed it.
Couldn’t reach the dashboard from the Mac. With Glances confirmed serving, the Mac still couldn’t connect. I worked outwards one layer at a time, checking that the container was running, the firewall port was open, curl to localhost was fine, ping from the Mac was fine. I then found out that macOS requires apps to be granted local-network permission, and Firefox had it off.
Local AI isn’t tenable. Ollama and Open WebUI install and talk to each other fine, but generation is painfully slow. With only a low-wattage laptop CPU and integrated graphics, inference falls back entirely to the CPU, and no dedicated GPU can be added to this board. A successful deployment experiment, but a failure as a daily tool.
What I Learned
- Working comfortably in the Linux command line and over SSH
- Docker basics: images, containers, port mapping, restart policies
- Why a firewall denies traffic by default and how you open ports deliberately
- A proper way to debug a networked service: test each layer in turn instead of guessing
- That reading hardware figures directly beats trusting the summary a system shows you
Future Plans
- Git server (Forgejo/Gitea) — host my own repos and learn Docker Compose
- Scraper — a scheduled job collecting data I care about
- Dashboard — one page linking the services together
Setup Reference
Exact commands and configs for every service are in the repo.