Setup, power and thermals, and software tips for running a Mac mini as a home server or self-hosting box.
The current home server setup uses an old Cudy router as a Wi-Fi access point and a Ubiquiti UniFi Cloud Gateway Ultra as the gateway. The main server has an i5-6500T processor and 8GB of memory, and it runs Nextcloud and Jellyfin with remote access through Tailscale. The second server has an Intel Pentium Duo processor and 4GB of memory, and it currently runs Immich. A QNAP NAS handles most network storage and backups. Possible next additions include Pi-hole or AdGuard Home for ad blocking across the whole home network, and Home Assistant for smart home automation. Another open question is how to split LXC containers and server resources between a busy main machine and a new empty machine. The rack itself is a low-budget 3D-printed setup, with cable cleanup and layout improvements still left to do.
A Vodafone internet setup in Ireland came with an Ultra Hub 7 router that changed behavior after an automatic update and reboot. Some settings were reset and then locked behind a message saying the router manages them automatically for better connectivity. The PPPoE details needed to connect through another router are already available. The main need is not many ports, because a separate switch is already in place. The required features are proper control over things like VLAN separation and speed limits. Building a custom router is not practical because time and budget are tight. Flint 2 is the current candidate because its web interface is familiar, and it still allows access through LuCI or SSH if deeper control is needed. The budget is 100 to 300 euros, with concern that Flint 2 may be overhyped or overpriced for this use.
A small mini PC that started with Home Assistant and Pi-hole grew into a self-hosted setup with a VLAN-separated network. The current setup uses a mix of n150 mini PCs and Lenovo m720q machines for Windows Server practice, domain controllers, SQL Server, and cyber security experiments. The firewall and router are still running in a temporary setup, with a plan to move later to a bare-metal OPNsense machine. Another m720q with four wired network ports is being considered because the machines offer strong value for the price. Storage runs on OpenMediaVault with a rough mix of 3.5-inch SATA drives and USB-attached drives, and that storage setup is the next likely upgrade. There is also a plan to build smaller racks by service type, such as one for networking and one for NAS storage. Grafana and Prometheus ran in Docker containers for about a year, but a custom asset manager and network overview app has been in development for the past few months.
A small home server still needs backups. The setup here is modest: Pi-hole, an MQTT broker, Grafana, InfluxDB, and Home Assistant, all likely fitting within a 500GB SSD. One option is to install Proxmox Backup Server on a separate HP ProDesk 400 G3 Mini and use it only for backups. That would work, but it may be more than this small setup needs. Backing up to an external SSD is also a realistic option if slow backup speed is acceptable. The most important part is not the size of the backup machine, but having regular backups and checking that restore actually works.
MacRumors’ expected roadmap says Apple may release about 20 new products by the end of 2027. On the Mac side, the Mac Studio, Mac mini, and iMac may move to M5, and the iMac may also get new colors. A larger Mac change could arrive in late 2026 or early 2027 with a MacBook Ultra using M6 Pro and M6 Max chips, an OLED screen, a touchscreen, and Dynamic Island. The iPhone lineup may include iPhone 18 Pro and Pro Max in September 2026 with an A20 Pro chip, a smaller Dynamic Island, and a variable aperture on at least one camera. A foldable iPhone Ultra may use a 7.7-inch inner screen, a 5.3-inch outer screen, and a Touch ID button instead of Face ID. iOS 27 may be adjusted for that foldable phone, including side-by-side apps like on an iPad. iPhone Air 2 may add an ultra-wide camera and longer battery life. A smart-home hub with its own screen is also listed.
Apple may ship a new entry-level MacBook Pro with an M6 chip as early as late 2026. The M6 could be Apple’s first chip made with a 2-nanometer process, moving on from the current 3-nanometer process. A smaller chip structure can fit more transistors into the same space, which can improve speed and power efficiency. A new design called WMCM would place the CPU, GPU, memory, and Neural Engine closer together so they can exchange data faster. Memory bandwidth could rise from about 153 GB per second on the M5 to about 200 GB per second on the M6. The GPU may also improve, with up to 12 cores, and the Neural Engine is expected to get stronger. These changes could matter for running AI tasks directly on a Mac. The M6 MacBook Pro is reportedly being tested first, while Mac mini, iMac, and MacBook Air versions may follow, but that part is still unclear.
Home Assistant can use Apple’s local AI model as a conversation agent in its voice setup. Apfel runs through a Homebrew service and exposes Apple’s local AI in a way that looks like an OpenAI-compatible server. Home Assistant then connects to it through Extended OpenAI Conversation. The useful role is general knowledge, not mainly smart-home control. It can answer questions about definitions, geography, history, and similar topics while someone is reading or working. After a few months of use, response times were fast and the setup was reliable enough for daily use. Setup instructions are available in the apfel-home-assistant GitHub repository.
A Mac Studio with an M1 Ultra chip and 128 gigabytes of memory became noticeably warm during a GPT-OSS-120b workload. The room felt warmer after only a few minutes. Power use reached a maximum of 153 watts while the work was running. When idle, the machine used about 10 watts and stayed cold to the touch. This firsthand experience felt much hotter than running several Mac minis and two MacBook Pros. The main point is that large local AI jobs can turn even an efficient Apple desktop into a real heat source.
A macOS laptop in the house recently had trouble connecting to a Synology DS720+ NAS, which raises concern about moving photo editing from a Windows tower to a future M5 Mac mini. The intended setup is to keep using three Synology NAS units for backups and shared mapped drives. Reports on Reddit describe macOS-to-Synology setups that can be flaky, including problems finding the NAS and reading or writing files. Lightroom cannot edit images directly from the network in the intended workflow, so the NAS would be for backup, not active editing. The DS720+ has extra memory and is the newest NAS in the setup. Storage capacity is not the issue, but needing to buy a newer NAS would add unwanted cost to the switch.
EleSync is a local memory vault for avoiding repeated setup details across Claude, ChatGPT, Gemini, and other AI tools. An AI tool that supports MCP can read from the vault and write new memories into it. The data stays in one folder on the user’s own machine as plain-text files, instead of being stored in a cloud service. Claude Desktop, Cursor, Cline, and Continue can connect through stdio with little setup. ChatGPT Developer Mode needs a remote-style connection instead of starting a local program, so EleSync also supports HTTP transport. The codebase is about 4,300 lines of Python and is released under the MIT License. Its five main tools are recall, remember, forget, memory_status, and find_conflicts. Storage uses files with YAML frontmatter plus content, while a local SQLite FTS5 index handles fast keyword search. Optional semantic search uses local ONNX embeddings on the CPU.
An old LG NAS has been put back on the home local network, and its password is known. The NAS appears in the Mac Finder server browser, but login fails when connecting to 192.168.7.87. The error says the server version is not supported. The server address shown by the Mac is afp://192.168.7.87. Changing the address to smb or FTP does not make the file connection work. The NAS web interface still opens normally in a browser through its IP address, so the device and network are at least partly working. The likely issue is a mismatch between the Mac and the old NAS file sharing method, not simply that the NAS is offline.
An old tower server, a NAS with six drives, a small switch, and a few Raspberry Pis can quietly raise the power bill when they run all day. After the bill rose noticeably over several months, each device was measured with smart plugs that track energy use. The old Dell tower used about 120 watts while idle, even though it was not doing anything critical. The NAS was not extreme by itself, but the always-spinning drives added more power use than expected. Lighter workloads were moved to a used Beelink mini PC, and the NAS drives were set to spin down on a schedule. The estimated savings came to about 18 to 22 dollars per month, depending on the month. That is not life-changing, but over a year it becomes real money for more useful gear. A UPS with monitoring could also track longer-term power trends for a whole rack.
Open Publica collects local city meetings, turns the audio into text, summarizes the content, and archives it. Most of the processing now runs on a headless M4 Mac mini. The machine has a 10-core chip and 24 GB of memory, and it processes about 110 to 150 meetings per day. A laptop is sometimes added to the worker pool when extra processing power is available. The processing system has been moved to a Temporal workflow setup, which helps make repeated, failed, or distributed jobs easier to handle safely. The transcription model is mlx-community/parakeet-tdt-0.6b-v3, chosen because it currently gives a good return for the workload. A free MCP endpoint is now available so an AI chatbot can ask about recent meetings, search the data set, and fetch full transcripts. The API is free, requires no login, and needs better documentation later.
FileFlows 26.06.9 is a monthly stable update with many new automation features for personal servers and home labs. The main addition is the Forge-powered FFmpeg Builder: Subtitle Resolver, which can organize subtitle workflows around target languages. It can remove unwanted subtitle tracks, convert image-based PGS/SUP subtitles into text-based SRT subtitles, translate existing tracks, and create new AI transcriptions from audio when needed. Custom Faster Whisper models are now supported for those audio transcriptions. A new FFmpeg Builder: Detect Language element helps workflows detect language and make better automatic choices. The Personal plan now allows 10 processing nodes instead of 5, so a home lab or personal server can run more processing work at the same time without extra cost. A new free FileFlows Database Tool can copy, back up, and restore a FileFlows database across SQLite, MySQL/MariaDB, and PostgreSQL.
Both home internet connections are stuck behind double NAT, so the home server cannot be reached directly from outside. The first goal is to run Pi-hole as a personal DNS server, then later use it for VPN access too. The problem is that dynamic DNS cannot find a public address that actually points back to the home network. Remote access also needs low delay because the server owner spends most of the time on another continent.
The setup would turn a Dell CAD workstation into a virtualization server. The hardware has an Intel Xeon CPU, 64 GB of memory, and an NVIDIA Quadro P2200 graphics card, and it currently runs Windows 11 Pro. The plan is to erase Windows 11, install Proxmox VE as the main system, and run several VMs on top of it. The planned VMs include Windows Server 2025 for Plex, a backup server, a local web server in a DMZ, and possibly a few lighter VMs later. The main concern is passing the Quadro P2200 directly into the Windows Server VM through GPU passthrough, so Plex can use hardware transcoding for H.264 and H.265 video. The open questions are whether free Proxmox supports GPU passthrough, whether the Quadro P2200 is stable in this setup, what common NVIDIA problems to expect, and whether Plex is better on Windows Server 2025 or Linux.
The current home server runs on a Raspberry Pi 4 with 4GB of memory. It hosts OMV, Plex with an *arr stack, immich, Home Assistant, and other self-hosted apps. The existing 2018 custom PC has a Ryzen 3 2200G, 16GB of DDR4 memory, a Corsair CX550 550W 80+ Bronze power supply, and an Asus PRIME A320M-K motherboard. The used HP G4 option has an i7-8700 and 16GB of DDR4 memory. Either machine should be a clear upgrade from the Raspberry Pi, especially if it can handle Plex transcoding. The main concern is power use, because the server will stay on all day and night; the Ryzen build seems likely to use more power than the HP G4, but that is not certain from the specs alone.
A new home homelab is being planned with Cat 6 cable running upstairs and at least one direct wired connection to the office. The Wi-Fi plan includes several wireless access points upstairs, 1 or 2 downstairs, and possibly one outside. The internet service will be Youfibre over FTTP, with 1Gbps upload and 1Gbps download. UniFi has been used before, but UniFi Express was disappointing, so the current choice is between Omada and MikroTik. The setup can handle more advanced networking because routing, firewalls, and address planning are already familiar. The needed choices include a router, wireless access points, switches, and the best overall brand direction. CCTV and a wired doorbell are also planned, probably using Reolink or Tapo cameras, with RTSP video saved locally. A basic homelab server will run a containerized NVR along with other local services such as HASS.
During a power outage, an Eaton 5SC UPS almost immediately began the shutdown process for its NUT clients. The UPS normally looked like it could run for about 30 minutes, but at the moment of the outage its display showed only 5 minutes left. The attached screens showed a sharp drop in both battery charge and estimated runtime when the blackout happened. The UPS was only about 2 years old, and it had seen only around two real outages in that time. Eaton’s ABM technology was expected to help preserve the battery. The battery self-test passed, and a later test by unplugging the UPS did not reproduce the sudden charge drop; it again showed about 30 minutes of expected runtime.
A small business owner wants to start using OpenClaw for work automation. The main goals are to reduce problems with running out of tokens while using Claude and GPT, and to keep work running overnight. Their current work laptop already has their files and work setup, but they do not want that laptop to slow down or carry the extra load. Buying a Mac mini is an option, but it is not clear how the Mac mini would connect to the existing laptop, files, website, search engine optimization work, LinkedIn posts, and other business tasks. The main questions are the easiest way to start, how to avoid major setup problems, whether the connections happen through OpenClaw or Claude, and how to use the setup well.
A 7-day measurement compared three homelab server setups: a Dell PowerEdge T360, a Minisforum MS-01, and an Intel NUC. The Dell PowerEdge T360 used about 104 watts on average. The power gap between a large tower server and a mini PC was bigger than many people might expect. That gap matters more when the machine stays on all day, every day, because the electricity cost keeps adding up.
A Synology DS220+ NAS is currently handling both storage and several server apps. It has 6 GB of memory and two 16 TB drives using Synology Hybrid RAID. Plex is running as a Synology package, not inside Docker. Several containers also run on the NAS, including Gluetun, qBittorrent, Sonarr, Radarr, Lidarr, Prowlarr, Dispatcharr, Mousehole, Calibre-web-automated, Shelfmark, Seerr, and Flaresolverr. The main choice is whether to add a mini PC for Plex and possibly the automation apps, upgrade the NAS memory, or buy a better Synology NAS. The mini PC would ideally run Ubuntu Linux, but setup guidance is needed because Linux is not familiar territory. A key question is how Plex on the mini PC would read folders stored on the NAS. A mini PC with an N150 chip is being considered.
The home has several TVs with different streaming limits, so watching something depends on which room is being used. One TV can use Google Cast for Jellyfin and Amazon Prime, another is the only one that allows Netflix sign-in, and a third is barely used because it only plays Amazon Prime. Small thin client PCs that were going to be recycled would be connected to each TV, so the TVs can act like simple screens. The goal is one central dashboard or controller that manages sign-ins for Jellyfin and streaming services. A web page or Android app would let someone choose a service and a movie or show, then send it to the chosen TV. Each thin client would connect to the controller and play whatever it is told to show on the screen.
A $290 1U firewall appliance ran pfSense for one year without stability problems. Its main parts include an Intel Alder Lake N100 processor, four Intel i226-V 2.5 gigabit network ports, and a PCIe x8 slot that can take two or four 10 gigabit SFP+ modules. It also supports M.2 NVMe, SATA, and Mini PCIe storage or expansion, plus HDMI and VGA for direct console access. The only fan is for the processor, so the machine stayed quiet. It also stayed cool in a home office without air conditioning and used little power. It worked with a wide range of SFP+ modules, but it may be more powerful than many small home labs need. For pfSense, a small fast SSD or NVMe drive is enough, and 2 GB of memory is enough, so buying 8 GB is unnecessary for this use.
A small home server setup needs to show a custom HTML maintenance page when one service is down. The setup uses Caddy with a reverse proxy and Cloudflare Tunnel. Changing Caddy error handling and adding failure timing to the reverse proxy has not solved it. The custom error page content can appear, but the HTTP header still returns a 502 status code. That lets Cloudflare treat the response as a server failure and show its own error handling instead. Upgrading Cloudflare for custom error pages is not a good fit because this is a free-plan home server for a few friends, not a business service.
Home server operators need a steady way to measure power use. The wanted setup is a device placed between the PSU and the wall socket. It should record current and voltage over time, not just show a one-time reading. The readings should be sent to an endpoint in a common format so other tools can store or display them. The price needs to fit a personal setup, not an expensive enterprise system.
A Mac running in clamshell mode is planned to use a new WD external hard drive as its main storage. The drive will hold media for a Plex server and will also store document backups and phone photos. The main concern is that the drive is not sold as a NAS-rated drive. NAS drives are commonly understood to be better suited for being powered on all day and handling frequent reads, and possibly writes. The practical question is how much extra failure risk comes from using a regular external hard drive 24/7.
NASSCAD 4 is a 3D CAD modeler that runs directly in a web browser with no install, account, or cloud service. The whole tool is a single 650 KB HTML file, so it can work fully offline. It supports 18 adjustable starter shapes, including generated screws and nuts that follow ISO and ASME standards. NassScript adds a JavaScript console for making models with repeatable instructions. It can export STL, OBJ, 3MF, GLB, and PLY files. The new feature is STEP AP242 B-Rep import and export, tested with Fusion 360 and FreeCAD. Practical compatibility with real-world STEP files is claimed at 99%. The tool is free under the CC BY-NC 4.0 license, which allows non-commercial use with attribution.
In California, PG&E peak power from 4 p.m. to 9 p.m. can cost almost three times as much as power around midnight. Selling extra daytime solar power to the grid for very little money, then buying costly power back in the evening, creates poor economics. A Tesla Powerwall would have cost about $16,000 installed, pushing the payback period beyond 9 years. A DIY setup using a 51.2V 100Ah Vatrer Power server rack battery and a 5kW Growatt hybrid inverter cost about $6,000, including an electrician’s final interconnect work and a critical loads subpanel. During the day, solar power charges the LiFePO4 server rack battery instead of being sent to the grid. From 4 p.m. to 9 p.m., the home runs key loads from the battery, including the fridge, chest freezer, router, lights, and devices. Those essential loads average about 2kWh during that window, leaving extra battery capacity.
A 1000VA GreenCell Powerproof UPS became very hot even with only about 30W of load. After 3 to 4 months, there was a burning smell and smoke, and one battery suffered an internal short that melted a hole in the battery. The UPS had been sold as a new open-box unit, but dust on it made old or previously used batteries seem possible. One proposed fix was to avoid putting new batteries back inside the hot case. The idea was to drill a side hole, keep the battery outside, make a 3D-printed bracket, use Powerpole connectors, and solder in a 220V-to-12VDC power supply plus a fan to move air through the existing vents. The later clarification was that the hole was in the battery, not the UPS body. The likely cause was an old battery, heat, or both, and replacing the batteries with a good-quality pair appears to be the more practical fix.