Self-Hosted
Self-hosted software put to a decision: measured comparisons, verdicts, and working setups — Docker managers, backups, auth, media, and AI.
Self-hosting in 2026 is not a hobby corner anymore — it is the default answer to subscription creep, and the hard part has moved. Installing an app takes one compose file; the real work is deciding which app, what it costs to keep running, and when the self-hosted route genuinely beats the service you are replacing.
About this section
That decision layer is what this section is for. The comparisons here get measured on real hardware where measurement is possible, and every verdict says who it is for and who it is not. Where a number matters — idle RAM, watts, break-even months — we publish the number and the method, not an adjective.
Start Here
Self-Hosted Services: What to Self-Host First
Six self-hosted services in the right order, with real hardware needs: Pi-hole, Uptime Kuma, Jellyfin, Nextcloud, Immich, and Home Assistant.
ToolSelf-Hosting Break-Even Calculator
Add your SaaS subscriptions, hardware budget, and electricity rate to see exactly when self-hosting pays for itself, month by month, over a 5-year projection.
The Decision Layer
Komodo vs Portainer vs Dockge: Best Docker Manager (2026)
Komodo for fleets, Portainer for Kubernetes and BE controls, Dockge for one Compose host. Portainer CE stays on 2.45 LTS as 3.0 goes Kubernetes-first.
HomelabCaddy vs Nginx Proxy Manager vs Traefik in 2026
The three homelab reverse proxies compared: Caddy, Nginx Proxy Manager, and Traefik on HTTPS automation, Docker discovery, licensing, and which to pick.
ComparisonRestic vs Borg vs Kopia: Which Backup Tool (2026)
restic vs Borg vs Kopia on verified 2026 docs: restic has had compression since v0.14.0, and where the backup lands decides the tool.
The Self-Hosted SSO Tax: A Verified Tracker (2026)
11 self-hostable apps that paywall SSO, each row verified against the vendor's own pricing page, plus free alternatives and fair-vs-rug-pull criteria.
HomelabBeszel vs Netdata vs Glances: Homelab Monitoring
Lightweight homelab monitoring compared: Beszel, Netdata, and Glances on footprint, history, and alerts — plus where Uptime Kuma fits, and which to run.
TutorialsSelf-Hosted Photo Library 2026: Immich vs PhotoPrism vs Ente
Immich vs PhotoPrism vs Ente in 2026: current versions, each project's hardware minimums, self-hosting and hosted costs, and how to get your library out.
Apps Worth Running
Paperless-ngx Docker Setup with OCR and Auto-Tagging (2026)
Full Paperless-ngx Docker Compose setup with PostgreSQL, Redis, Tika, and Gotenberg. OCR every document, auto-tag by rule, and search your whole paper history.
TutorialNextcloud AIO Mastercontainer: Docker Setup Guide (2026)
Deploy Nextcloud All-in-One in Docker: Postgres, Redis, Nextcloud Office, Talk, and auto backup. Replace Dropbox, Google Drive, and Docs with a stack you own.
TutorialUptime Kuma Docker Setup, and Where to Actually Run It (2026)
Uptime Kuma in Docker Compose, plus the part most guides skip: a monitor that lives inside the network it watches cannot tell you that network went down.
AI in the Stack
Ollama vs llama.cpp on an RTX 5080: Measured (2026)
Same GPU, both servers benched: llama.cpp decodes 2-6% faster (more on gpt-oss, via different conversions), Ollama prefills 4-7% faster. One model won't port.
TutorialOllama in Docker with GPU: Stop Installing CUDA Inside WSL2 (2026)
Ollama with NVIDIA acceleration in Docker, on native Linux and WSL2, plus Open WebUI. Why the CUDA install everyone does inside WSL is unnecessary.
TutorialsHow to Self-Host a Local LLM on a Single RTX 5060 in 2026
Run a private LLM on a $300 RTX 5060. VRAM tradeoffs, software stack picks, Windows + Linux setup, real tokens/sec numbers, and a verdict.
DatasetRTX 5080 Local LLM Throughput: Measured Tokens/Second Dataset
Measured decode and prefill tokens/second for gpt-oss 20B, Qwen 2.5 14B and 7B, and Llama 3.2 3B on a retail RTX 5080, with VRAM and load times. CC BY 4.0.
More from this section
Self-Hosted AI Agent: What It Is and What I Run (2026)
A self-hosted AI agent is the loop, tools and memory on your hardware; the model is a separate choice. The 2026 runtimes, three configurations, and what I run.
LM Studio vs Ollama (2026): Measured Speed and RAM
LM Studio vs Ollama, measured on one RTX 5080 with the same GGUF files: Ollama wins dense decode, LM Studio wins prefill. Pick by workflow; fix your settings.
Best MCP Servers for a Homelab: 11 Measured (2026)
Eleven MCP servers measured on real hardware: 178-422MB RAM each, ~1GB for a five-server stack. Which earn their memory, which to skip, working configs.