Setup, power and thermals, and software tips for running a Mac mini as a home server or self-hosting box.
A large Mac restore job needs to erase and reinstall a few hundred Intel MacBook Pros from 2017 to 2020. The current setup uses Two Canoes MDS to create a disk image with the macOS installer, puts that image on a web server, and mounts it from each MacBook before running the installer. The web server is the built-in Apache server on macOS. The problem is that this is unicast, so the server sends a separate copy of the same data to every MacBook. For restoring many computers at once, multicast may be better because it can send the same data to many machines at the same time. The needed missing piece is a tool that can mount a disk image in recovery mode using multicast.
CT ROADMAP is a self-hosted visualization tool for mapping home lab hardware and the jobs each part performs. Home server setups can become hard to understand once custom flows, scripts, config files, and special run commands start to pile up. The tool is meant to show both the physical layout and the small pieces that make the system work. It is available as a Docker image. It does not scan the network by default; everything is added manually with drag and drop. It can also be used to plan future features or settings before putting them into service. Possible future additions include light network scanning for the first setup, clickable links for URLs and file paths on tiles, and other requested features.
A family photo-sharing cloud is being run from home. Nextcloud is hosted on a Pi5, and access is limited to devices using Tailscale. The problem is that turning on Tailscale before using Nextcloud feels too complicated for a family member, so the service becomes hard to use in practice. Keeping the setup private protects the home network, but making it easy for the family may pressure the operator to weaken that protection. The risk is worse because the family member reuses nearly the same password across many accounts. The setup also needs to work when that person is at work, without the server operator present to help.
A cheap network cabinet is being used as the base for a home lab and home server setup. The plan also includes running Cat6 cables through the home so devices can use wired internet instead of relying only on Wi-Fi. An old PC from work will be the first server machine. The main question is where to begin learning. Proxmox and TrueNAS are named as possible starting points.
A small home server rack setup is continuing to expand. Most of the equipment is connected over 10Gb Ethernet, using both fiber and copper links, while some parts still use 2.5Gb Ethernet. A Ugreen DXP2800GT storage unit provides shared storage to Proxmox through NFS and iSCSI. Some SFP modules may be moved to reduce strain on the cables.
A home server in Beaverton, Oregon runs the whole flow locally, from payment to publishing. The setup uses StartOS, a full Bitcoin node, LND, Alby Hub, Ollama with the phi3:mini model, and Ghost for the blog. A visitor goes to blog.newellfamily.org/store, pays 1 sat over Lightning, and enters any topic without creating an account or giving an email address. A small AI model running on hardware in the house writes a blog post and publishes it live in about 90 seconds. Basic safety checks are included: a keyword blocklist and a model check of the input before writing. There is also a paid “How I Built This” writeup for 2,000 sats and a 500-sat tip jar for testing a Lightning payment. The hardest practical problem was inbound liquidity, which is the ability to receive Lightning payments. Megalith LSP through Alby Hub made the first payment clear in under a second, and the 1-sat price was mainly a proof that the whole payment path worked, not a real revenue plan.
For development use, Mac mini memory can matter more than storage. A 16GB model may be tempting, but 24GB or 32GB can feel like a safer minimum for longer use. Large storage is less important when the machine is not meant to run local AI models. Apple refurbished stock alerts can help catch a specific configuration before it sells out. A 24GB memory and 256GB storage model became available, was ordered quickly with Apple Pay, and arrived shortly before prices increased. The practical lesson is to buy the useful configuration when the price works, instead of waiting endlessly for a possible M5 model and risking higher prices.
Two old Pixel 3a phones were reused instead of being thrown away. PostmarketOS was installed on both phones, then K3s was installed to create a small Kubernetes cluster. Some network settings had to be fixed first because the phones could not reach the internet. A used OnePlus 6T was later bought for €50 and added as a worker device, making the current setup 3 devices in total, including one control device. The setup is slower than a mini PC or a proper server, but it is enough to run 3 Hermes agents and a full Grafana setup. All devices connect to the home network over Wi-Fi. Two more phones are waiting to be added: a Poco X3 Pro and a Pixel 6a, each bought for €50. The Poco X3 Pro should join soon, but the Pixel 6a cannot be used yet because PostmarketOS does not recognize its Wi-Fi chipset.
The main router is in the living room, and the house layout makes it hard to run CAT6 cable through the walls. A router running OpenWRT is already being used as a wireless bridge so a PC on the other side of the house can get internet access. Two server machines currently stay in the living room and plug directly into the main router, but the goal is to move them elsewhere. The practical question is whether the wireless bridge can feed a network switch, with the servers and other devices plugged into that switch. Wired networking is understood to be better, and bandwidth may be limited, but the main concern is whether this setup can work at all.
Coolify is being used as the main control tool for a self-hosted setup. It was chosen to automate the reverse proxy and SSL certificate work, and it runs on Oracle’s free cloud tier. The services are split into two groups. Core infrastructure currently includes Authentik and Headscale, while public web services include a club introduction site, a portfolio, a journal, and more sites to be added later. The Oracle instance had to be recreated twice because of setup mistakes and one termination from Oracle’s side. Coolify and Authentik are now set up, and the next steps are setting up Headscale, connecting OPNsense through a Tailscale client, then adding Navidrome and Nextcloud.
The topic is backup planning for a Proxmox homelab server. The available item data does not show the server setup, storage devices, backup targets, restore method, or budget. The main issue is how a home server owner can keep data and settings safe if the machine fails, storage breaks, or the server setup becomes unusable.
The setup began as a lab for Network+ study and hands-on practice. It used two MacBooks and a used Dell OptiPlex Micro. The MacBooks ran Windows 11, Ubuntu Server 24.04, Windows Server 2022, AD DS, DNS, and IIS for practice. The Dell OptiPlex Micro was the machine that changed roles. It started as an always-on computer for remote access, breaking things, and rebuilding them. Over time, it became the small server that handled real tasks. It now runs a few Python automation programs through Task Scheduler. One handles repeated SEO work for a study site, and another sends weekly analytics emails for iOS apps. One automation stopped working for four weeks with no obvious crash or error. The missing weekly emails were not noticed for a month because nothing actively warned that the job had gone quiet.
Fort is a CLI tool for checking and fixing security settings on a Mac. It runs from the terminal instead of a normal windowed app. The available information points to one main purpose: review the Mac’s current security setup and help correct settings that are not in the desired state.
Tailscale is a tool for making VPN connections between personal devices easier to manage. A home network that already has WireGuard and OpenVPN can already handle basic remote access, such as checking a home server while away or using a safer connection on untrusted Wi-Fi. Tailscale’s main difference is that it handles key management, changing internet addresses, and NAT traversal for devices behind routers. That is especially useful when many devices or locations need to connect directly to each other. For a simple setup where the main goal is reaching one home network, a self-hosted WireGuard or OpenVPN setup may already solve the problem. The real question is whether the home lab needs easier management across many moving devices, not whether Tailscale is automatically better.
FileBeam is a file transfer tool that runs from the Mac terminal. When it starts, it opens a local server on the Mac and shows a QR code. An Android phone can scan the QR code, open the page in a browser, and transfer files with the Mac. No phone app is needed. It is aimed at Mac and Android transfers, but iPhone and iPad can also use it. It works over the same local Wi-Fi network and does not need cloud storage or the public internet. It shows live transfer speed, progress, and estimated time left. Transfers can be canceled, and cleanup is handled automatically. The code is available as open source on GitHub.
The setup uses a MacBook Pro M4 Pro with 48GB of memory and a Mac mini M4 with 16GB of memory to build a personal AI assistant and knowledge system. The goal is not just to run a local LLM, but to create a persistent “second brain” that remembers work, health, projects, documents, and personal knowledge over time. The MacBook would run Hermes, a local Qwen model, browser and computer automation, voice and chat controls, and the main decision-making layer. The Mac mini would stay on as the backend for long-term memory, document indexing, a vector database, embedding generation, background summaries, MCP tool servers, and nightly maintenance such as re-indexing, deduplication, summaries, and backups. The knowledge base would include PDFs, emails, notes, drawings, and similar personal material. The proposed knowledge structure follows Andrej Karpathy’s LLM-WIKI idea inside an Obsidian vault, with fixed source documents, AI-maintained Markdown pages, an index page, and Obsidian wiki links connecting everything. The vector database would retrieve relevant information, while the Obsidian wiki would become the readable and maintained knowledge layer. The two Macs would be linked with a CalDigit TS3 Plus dock, an OWC Thunderbolt 5 cable, and Thunderbolt Networking.
A base M4 Mac mini with 16GB of memory and 512GB of storage is fast enough for this setup, but the memory may be too small. The processor is rarely using more than 10%, so the extra speed of an M4 Pro is not needed. The real target is a base M4 Mac mini with 32GB of memory and 512GB of storage, but that new configuration appears to be unavailable from Apple. The available upgrade path seems to push buyers to an M4 Pro Mac mini with a 12-core processor and 48GB of memory. That costs about $1,000 more than the wanted base M4 32GB setup. A used M4 Mac mini with 32GB of memory and 512GB of storage may be available for about $1,100 to $1,200. The practical choice is between buying that used model now or waiting for a future base M5 Mac mini.
The claim that RAM and SSD prices will stay high until 2030 or 2033 is stronger than the evidence supports. The clearer point is that Micron has made long-term memory supply deals with large customers. Those deals help big companies reserve future supply and make their costs easier to plan. They do not prove that regular buyers will face fixed high prices for every RAM or SSD purchase. Home server buyers are treating this as a reason to keep current hardware longer, watch used parts, and avoid wasteful upgrades. Some expect AI server demand to keep pressure on memory and storage prices, while others think new supply from Chinese memory makers such as CXMT could change the market later.
An M4 Mac mini was running Home Assistant OS as a virtual machine in VirtualBox. The setup used the `haos_generic-aarch64-17.2` image, which appeared as ARM 64-bit Oracle Linux and ran normally on the ARM-based Mac. Because the Home Assistant OS problems seemed tied to VirtualBox rather than Home Assistant itself, the same virtual machine was exported from VirtualBox as an `.ova` file and imported into VMware Fusion Pro 26H1. The import appeared to work. When started, VMware Fusion refused to power it on and said the virtual machine requires the X86 machine architecture, which does not match the ARM host. The confusing part is that the original image was ARM 64-bit, but Fusion treated the imported machine as if it needed X86.
On June 25, 2026, Apple raised prices for almost its whole hardware lineup in Germany. The affected products include Macs, iPads, HomePod, HomePod mini, Apple TV, and Vision Pro. The Apple TV 4K with Ethernet rose from 189 euros to 299 euros, a jump of about 58%. The entry iPad with A16 rose from 379 euros to 499 euros, and the iPad mini rose from 549 euros to 679 euros. The 13-inch MacBook Air now starts at 1,399 euros instead of 1,199 euros. The HomePod mini rose from 109 euros to 139 euros, and Vision Pro now starts at 3,999 euros instead of 3,699 euros. iPhone, Apple Watch, and AirPods were not included for now. Apple points to a memory chip shortage caused by rising demand from new AI data centers. Tim Cook had already warned in mid-June that price increases had become unavoidable and that the situation was no longer sustainable. Microsoft raised Xbox prices on the same day, and Dell, Lenovo, and HP have also raised prices because of the memory shortage.
The home server has been upgraded to a much more powerful desktop-class setup. The new system uses a Threadripper 9970X, 128GB of DDR5 memory, an RTX Pro 6000 graphics card, a 2TB M.2 drive, about 48TB of ZFS storage, and two 10GbE network cards. It runs Ubuntu 26.04 LTS. The motherboard needed a BIOS update before it could recognize the Threadripper 9970X, which was expected but still stressful. The machine is now being used to train a draft model for speculative decoding on Qwen 3.6 27B. The older 9900K system is planned to return later as a Proxmox node.
A used Dell OptiPlex 7060 Micro was turned into a small home server with two 6TB external hard drives. The computer has an Intel Core i5-8500T processor, 16GB of memory, and a 256GB NVMe SSD, and it cost $150. The two Seagate Expansion drives cost $280, and the power cord cost $34, bringing the total to $464 for 12TB of storage. The server runs several services through Docker: Immich for photo and video backups, Jellyfin for media, Navidrome for music, Pi-hole for ad blocking, Tailscale for remote access, Uptime Kuma for uptime checks, Homepage Dashboard, Ollama with Open WebUI, and Kopia for backups. One 6TB drive is the main storage drive, and the second 6TB drive is used as its backup. The goal is to reduce paid cloud subscriptions while learning Linux, Docker, networking, and self-hosting. Immich may also be shared with a few close friends so each person has a private photo backup space.
A Comcast modem was replaced with a self-owned UniFi modem in a rack, and the internet connection worked normally for several weeks afterward. The modem was activated through the Xfinity app by entering its MAC address and completing the setup flow. Speed and latency were satisfactory, but the Xfinity app and portal kept showing the equipment as offline. Repeating activation and running the troubleshooting steps did not fix the status. Some sites, especially remote work servers, sometimes returned an HTML response saying the Xfinity internet service was not activated. The same offline status appeared in the Xfinity portal from a phone outside the home network, so the issue appears to be a mismatch inside Xfinity’s system rather than a local device or malware problem.
An Odroid N2+ is currently being used as a headless Plexamp playback device, while the Plex server runs on a Synology NAS. The NAS is too old to support Container Manager, so it cannot easily run newer container-based services. An older Intel Mac mini is being considered as a machine for Home Assistant. The main concern is that running Home Assistant in a Docker container may not provide full functionality. The practical questions are whether one Mac mini can run both Plexamp and Home Assistant, whether full Home Assistant functionality requires a virtual machine, and whether Plexamp playback can be added after installing Home Assistant OS on the Mac mini.
An M4 MacBook Air is running into rendering bottlenecks during game-making work. Its fanless design makes it less suitable for long, heavy rendering jobs. The next machine could be an M4 Pro Mac Mini, a Framework Desktop with an AMD Max+ 395 chip, or an M4 Max Mac Studio. The Framework Desktop would cost about $600 more than the M4 Pro Mac Mini, while the M4 Max Mac Studio would cost about $900 more. The Framework Desktop is appealing because it can run Linux, especially OpenSUSE. It could also be used for local LLM work alongside a home assistant setup. Blender Open Data has almost no useful AMD Max+ 395 results, apart from some CPU render scores that do not answer the graphics performance question.
A new Mac Mini M4 with 24GB of memory is being set up with the goal of using a local LLM instead of relying on outside AI services. The main concern is whether this hardware will feel too slow or produce weak answers. The practical question is which model works best on a similar Mac Mini setup. There is also concern that full local use may not be enough, and that a cheaper external Chinese model may still be needed for useful results.
A Plex server works at home but cannot be reached from outside the local network. The server runs on a Mac Studio. Some media files are stored on the Mac Studio, and some are stored on a newly bought NAS with two hard drives. The NAS is wired into the network and has its own IP address. The Mac Studio connects over Wi-Fi. The internet service is AT&T Fiber, with the AT&T-provided modem’s Wi-Fi turned off and IP Passthrough also turned off. A Google Nest Pro mesh Wi-Fi system sits behind the modem, then connects to a gigabit switch, and the switch connects to the NAS. In Google Nest settings, UPnP is on, the WAN connection uses DHCP, and the Mac Studio has a reserved static IP address.
In India, the base M4 Mac mini with 16GB memory and 256GB storage became harder to buy and more expensive. One recent price comparison showed the same machine costing about ₹35,000 more than it did two months earlier. The same configuration was unavailable or out of stock across online platforms, while some returned listings appeared at a higher price. Refurbished base units also appeared briefly and disappeared quickly, which suggests cheaper stock is being bought fast. Some orders placed soon after a price increase were also cancelled. A common expectation is that a real price drop may not arrive until a newer Mac mini generation, such as M5 or later, replaces the current model.
The plan is to protect home server services such as Home Assistant and Jellyfin. Local network access would use SSL certificates issued by AD CS on Windows Server. Public internet access would go through Nginx Proxy Manager with Let's Encrypt certificates. The main question is whether separating local and public certificate handling this way is a good setup.
A 4-bay external drive enclosure is needed for four 3.5-inch hard drives. It will be connected to a computer that is being used as a server. The storage is intended to run as RAID. The main uses are Jellyfin for storing and playing media, plus server storage for photos and documents. The main problem is finding an enclosure that is reliable enough for this kind of always-on use.