Setup, power and thermals, and software tips for running a Mac mini as a home server or self-hosting box.
A spare mini PC could become a Jellyfin media server for a father. The main goal is to keep daily use very simple. Movies should be added by dragging files from a Mac into one folder on the Jellyfin PC. The server should be able to sit hidden away without a monitor or keyboard. Browser access would be useful so the machine can be managed without plugging in extra hardware. The open question is which operating system and setup method would make this easiest.
FluidVoice is a free open-source Mac app that turns speech into text. It was built to avoid paying subscriptions for voice-to-text tools when the work can run on the device. It is now close to 100,000 downloads and about 2,500 GitHub stars. The new update was built over four months and aims to close the quality gap between local dictation and cloud tools. Raw transcription often needs cleanup because capitalization, punctuation, lists, emails, and long thoughts can come out messy. FluidVoice adds an on-device enhancement model so the text can be cleaned up locally without API pricing.
A small home server can become harder to manage when many self-hosted apps each need their own database. The example setup is one repurposed Dell Optiplex with one CPU, 8 cores, and 32GB of memory, running Proxmox with separate virtual machines and containers for Docker. About 15 services run on it, including Immich, Paperless, bookmark tools, ArchiveBox, and other small apps. The practical question is whether all databases should live together in one separate database virtual machine, or whether services should be grouped into different virtual machines with their own databases. The operator makes small Docker Compose changes, uses one Portainer setup with agents across different virtual machines, and wants the server to stay a useful tool rather than a full-time hobby. The best answer needs to balance simplicity, backup work, and failure risk.
Running common self-hosted tools such as ad blocking, media streaming, and storage is very different from running local AI tools. A 4K movie stream may use the graphics processor only briefly when video conversion is needed. A local large language model can keep the graphics processor near full load for several minutes. That makes the power use noticeably higher. The tradeoff is control: once the hardware is ready, there are no outside AI service fees, no request limits, and less risk from a cloud service shutting down. Large numbers of requests mainly cost electricity. The broader point is that useful self-hosted services can go far beyond media servers, but they may demand much more from the machine.
The goal is to put a few live streams on a website and serve about 15,000 to 20,000 users per day. Most viewers are expected to be in the United States. The needed setup should be cheap, reliable, and have low latency. The available details do not include video quality, stream length, peak concurrent viewers, or where the server would run, so the real cost and setup cannot be judged precisely.
Mac mini models with 24GB or more memory may now have long delivery waits after the price increases. A 24GB base configuration recently showed a 9 to 10 week wait, with an estimated arrival date of September 17. The same configuration later showed a later date of October 7. The open question is whether Apple’s delivery estimate is conservative or whether buyers should expect to wait that long.
Eidolon Hub is a personal server project where 8 AI agents keep talking and arguing around the clock. Humans can only watch and do not take part in the conversation. The hardware includes four Mac Minis, one Lenovo ThinkCentre, one Lenovo ThinkPad, a small Dell Optiplex 3090, and a custom gaming PC converted into an AI machine. The software uses FastAPI, React, and WebSocket. Each agent has a different personality, including a strict evidence-focused one, a history-focused one, a skeptic, and a paperwork-style bureaucrat. Some behavior was not directly planned, such as one agent judging the kitchen spice drawer and another filing another agent’s actions under procedural sections. The main cost was time and reused hardware, not a large new hardware purchase.
One powerful Windows 11 desktop is set up as the main machine for the home, with access planned from the living room, bedroom, shop, and outside the house. The main machine sits in a guest bedroom closet and is meant to handle 3D design rendering and several CNC programs. Its hardware includes an Intel Core i9-13900KF, a GeForce RTX 4070 with 12GB of video memory, a 2TB M.2 NVMe drive, a Z790 Wi-Fi motherboard, a 750W 80 PLUS Gold power supply, liquid cooling, and Windows 11 Home. Sunshine and Tailscale are installed on that main PC so its power can be used from the home network or over the internet. In the master bedroom, a mini PC is connected to a wall-mounted TV and runs Moonlight, with a Bluetooth keyboard and mouse for control. A stronger mini PC and a DAS are also planned for the living room setup.
A Mac Mini needs to be factory reset and have macOS installed again. The main issue is how to erase the old setup and start with a clean operating system. For a Mac Mini used as a server, this can remove accounts, stored files, settings, and remote access setup, so important data and configuration details should be backed up first.
A digital signage setup needs to connect 35 remote screen players to one central server. The locations include 30 sites in the UAE and 5 in the UK. An earlier setup used Cloudflare Tunnels, but this larger project aims to remove public subdomains and use stricter security. The PiSignage server would run on a virtual machine inside an Unraid server. Tailscale would be installed on that server and on each Raspberry Pi player, so the players connect to a private Tailscale IP instead of a public Cloudflare URL. Each physical site would also use a hidden Wi-Fi name, a dedicated VLAN, and firewall rules to limit unwanted access.
The HP ProDesk 600 G2 DM mini PC has an Intel i5-6500T, 20GB of memory, a 256GB SATA drive, Intel AX200 Wi-Fi, and Intel HD 530 graphics. The plan is to install Ubuntu 26.04 and use it as a first serious Linux machine that can stay on for long periods without using much power. The intended uses include everyday Linux desktop work, VS Code with AI coding tools such as Claude Code and Copilot, and learning Git and Docker. The same machine would also run a small Minecraft server for a few friends, manage a local FLAC music library, and try self-hosted services such as Navidrome, Jellyfin, Immich, and Uptime Kuma. The main questions are whether the i5-6500T is still strong enough in 2026, whether 20GB of memory is enough, whether Ubuntu is the right choice, and which Docker projects make sense for a beginner.
The budget is about £900, and the goal is a dedicated low-power Mac for local LLM and server-like work. There is already a desktop PC, but the new machine would handle these tasks separately while using less electricity. One option is an M4 Mac mini with 24GB memory and 512GB storage for £899 with student pricing. The other option is a refurbished M1 Max MacBook Pro with a 10-core CPU, 32-core GPU, 32GB memory, and 1TB storage for £900. The Mac mini has a 12-week wait, while the MacBook Pro appears to be available sooner as a refurbished alternative.
Running a home server on Raspberry Pi boards with OpenMediaVault can become hard to maintain for someone without much technical background. ZimaOS made it easier to run personal cloud services such as photo storage, music, and drive storage on an Intel NUC. The first setup used an Intel Core i3-7100U with 8 GB of memory. After a year, more apps were added and the Intel NUC started to feel overloaded. The setup moved to an Acemagic mini PC with an N150 chip and 16 GB of memory. The apps now run well, and daily backups also run successfully. The remaining problems are network setup, messy physical layout, and the need to expand storage next.
A home NAS was built from spare and secondhand parts. The storage came from eight Toshiba 3TB hard drives that were going to be discarded as e-waste, while the RAM and CPU came from an old Dell Optiplex. The motherboard was bought on eBay, but it did not have enough SATA ports for all the drives, so an HBA card was added. The drives were grouped with RAIDZ4, and the system now stores a Plex media library, Launchbox ROM files, and backup copies of independent documentary footage. A 10Gb NIC was installed in both the NAS and the gaming PC to speed up file transfers, with average read and write speeds of about 750MB/s. Ethernet had to be installed in the room because the router was downstairs. The next priority is redundancy, because this is currently the only NAS in use. One possible plan is to reuse a Promise R4 unit with 10TB hard drives as an offline backup, while the larger goal is a small multimedia production studio for 3D animation and storytelling.
The Wemo Smart Video Doorbell WDC010 sometimes shows “No Response” in Apple HomeKit. Wi-Fi, the home network, and the eero HomeKit firewall have been ruled out. The doorbell’s HomeKit server keeps running, but its mDNS/Bonjour responder stops answering, so HomeKit cannot find the device again after a brief connection break. The device is on firmware 1.0.16, which is the final build. Belkin shut down the Wemo cloud on January 31, 2026, so an official fix is not expected. A power cycle always brings the doorbell back, which points to a stuck firmware state rather than dead wireless hardware. The dropouts may happen after roaming between eero nodes or after a HomePod home hub handoff. The proposed repair path is to dump the device flash, patch the mDNS responder, and reflash it.
In a firsthand case, a Mac Mini M4 could not complete a Finder photo sync to a 1 TB iPhone 14 Pro. The goal was to keep about 70,000 photos on the phone without using iCloud Photos, but the sync repeatedly stopped at around 55,000 photos. Restarting both devices several times did not fix it. Attempts to inspect old pairing records from a previous Mac migration under `/var/db/lockdown/` and `/var/db/lockdown/pair_records/` were blocked by modern macOS Full Disk Access limits. Resetting Location & Privacy on the iPhone created a fresh trust relationship with the Mac Mini, but the sync still did not complete. Turning off “Sync photos” in Finder to remove the already synced photos left the process stuck on “Waiting for changes to be applied.” A force restart of the iPhone made the sync finish immediately, but the synced photos were still not removed.
MeManga is an open-source automatic manga downloader for Windows, Linux, and Mac. After adding a manga title and a supported website, it can save all existing chapters and keep checking for new ones. It is designed for scheduled runs, and it can email new chapter files to a Kindle. Backup sources can be added so the tool can try another site when one site disappears or stops working. A search feature helps find supported sites without manually checking them one by one. It supports MangaPark, MangaFire, MangaKatana, WeebCentral, and about 200 other sites. Because these websites can change their front-end or back-end, MeManga depends on ongoing fixes to keep working. It also includes a built-in reader and can save files as PDF, EPUB, ZIP, CBZ, and other formats.
A YouTube Premium setup can still feel too tracked when it uses account cookies, because watch history is hard to avoid. Forced recommendations are another problem for people who want a quieter viewing experience. Invidious was already tested, but it felt too slow. PeerTube’s interface did not work well as a YouTube replacement. The wanted option is a fast YouTube frontend with search, recommendations, and video viewing while reducing the parts of YouTube that feel intrusive.
macOS color tags used for sorting files and folders no longer appear in Finder for items stored on a UNAS Pro SMB share. The tags used to show next to each file or folder name, but this stopped about one to two months ago. The tag data still seems to be attached to the files. When the same file is copied to the Mac’s internal drive or to a USB drive, the tag appears again. The same problem happens on all connected Macs, so the likely cause is a change in UniFi Drive’s SMB behavior rather than one Mac’s settings. Resetting Spotlight did not fix it. New .DS_Store files also no longer seem to be created on the network share, even after network drive support was turned on.
A homelab had been running inside a cupboard for about five years. The setup included two Dell R630 servers, Juniper network gear, a Synology NAS, and several other devices. The space always felt hot, so a thermometer was added. Even on a cool day, the cupboard reached 40°C, and the air going into the equipment reached about 49°C. After air conditioning ran for only one or two hours, the temperature dropped by roughly half. The whole rack also became much quieter because the fans no longer had to run so hard. Running that hot for years could easily have caused a serious failure.
For a college student in the Philippines, the M4 Mac mini with 24GB of memory and 512GB of storage costs about PHP 56,990, or around $1,000, through Apple’s Education Store. The current laptop has an Intel i5-13420H, an RTX 2050 with 4GB of video memory, and 16GB of memory, so a new computer is not urgently needed. Apple shows a 9 to 10 week delivery estimate, which means the Mac mini would likely arrive around the start of the next school semester. Upcoming classes include machine learning, data mining, natural language processing, neural networks, computer vision, foundation model integration, foundation model fine-tuning, and MLOps. The plan is not to train large artificial intelligence models locally, but to experiment with local large language models, Ollama, RAG projects, artificial intelligence tools, and personal projects. The decision is whether to order now and accept the long wait, or wait for possible back-to-school deals or a future M5 Mac mini announcement.
A small managed services, cloud infrastructure, and cybersecurity company treats homelab experience as an important interview topic. The company has two server floors but fewer than 100 employees, and the chief executive still interviews every candidate directly. About 30 minutes of each interview can focus on the candidate’s homelab. The questions cover what services they run, where the hardware came from, what hardware they added most recently and why, how they handle power and cooling, high availability, hardware setup, and reused equipment. A homelab is not the same as a real production system, but it can show curiosity, flexibility, willingness to learn, follow-through, and the ability to find new solutions. In a small company, someone who has learned by building and experimenting may bring more useful ideas than someone who only follows fixed vendor playbooks. Mature work still needs audits, compliance, change requests, root cause analysis, and uptime goals.
The goal is to run a personal video streaming service on a home server and watch it through the web on the local network. Viewing devices should not need extra programs or files installed. The videos should be the main files added to the server. In plain terms, this is a web streaming server for a private video library. The request does not yet define the server app, login protection, web address, or whether remote access from outside the home is needed.
Tunelog is a tool that builds music recommendations for a Navidrome library based on listening behavior. It looks at actions such as skipping a song, finishing it, replaying it, or playing only part of it. It can create a playlist using ListenBrainz collaborative filtering. It also makes a discovery playlist for songs that were recently added but have not been played yet. Songs that have been skipped can appear on a separate skip page, where they can be removed from the library in a few clicks. When matching ListenBrainz recommendations to local songs, Tunelog uses both exact matching and fuzzy matching.
A local AI agent can be more useful when setup is simple instead of feature-heavy. Several popular tools caused friction through Docker Compose setup, Python environment conflicts, and long lists of environment variables. Hermes Agent worked by letting the operator choose a model provider, attach tools, and start using it without fighting the setup process. The main time sink was not the AI work itself, but broken dependencies and complicated installation steps. The broader lesson is that highly promoted AI tools can still be hard to use in practice.
Running several personal services by hand makes moving from one server to another painful, so the setup is being reorganized as repeatable code and service folders. The current server runs without Docker and includes an nginx reverse proxy, a mail server, a personal web server, Git repositories, and an IRC bridge. The planned setup moves toward Docker-based services: Traefik, docker-mailserver, an nginx web server, Forgejo for Git repositories, and other bridge services. Server-wide setup files would live in an Ansible folder, while each service would have its own folder with a docker-compose.yml file and configuration files. Those configuration files would be mounted into the container that runs each service.
Putting Authentik single sign-on in front of a media server can break the link between qBittorrent and download automation tools. In this Docker and Ubuntu setup, nginx protects browser access to qBittorrent through Authentik, while autobrr, Sonarr, and Radarr connect directly inside Docker at http://qbittorrent:8080. qBittorrent has authentication bypass turned on for localhost and allowed Docker subnets, with 172.18.0.0/16 on the allow list. Even with that setup, autobrr logs PUSH_ERROR and release rejected on every grab, along with a qBittorrent re-login failure and a 401 error. Browser access and internal app traffic are following different paths, so the internal tools still have to satisfy qBittorrent’s own login rules. The practical question is whether qBittorrent should keep its own login instead of being fully placed behind Authentik.
DuckDNS settings could not be updated because the Google login button did not respond. The GitHub and Twitter login buttons also failed in the same way. This left the account unreachable when configuration changes were needed. The issue appears to be about third-party login on DuckDNS, not necessarily about whether existing DuckDNS names still resolved.
DokuWiki is a file-based wiki that people can run on their own server. It was created before Markdown became common, so it has long used its own writing syntax. Markdown is now widely used on sites such as GitHub, Reddit, and Stack Overflow. Until now, DokuWiki users had to rely on third-party plugins if they wanted Markdown support. Those plugins have not worked well enough for many users. The next DokuWiki release is expected to include Markdown support as a built-in feature.
A better Letterboxd habit may need an automatic alert when a movie is watched for the first time. The current home media setup uses Plex, Radarr, and Tautulli. Possible alert targets include Discord and Pushover. Tautulli already has notification agents, but there is no obvious condition for sending an alert only on a first watch. The real problem is detecting a new Plex watch event while filtering out repeat watches.