Self-Hosted Apps: What's Actually Worth It

Every "self-hosted apps" roundup is the same unvetted list of app names — Immich, Nextcloud, Vaultwarden — with no first-hand testing and no honest "skip this one." The real ceiling on what's worth self-hosting isn't the software. It's whether you'll actually maintain it in six months, and that depends entirely on which category you're replacing.

TL;DR

A mini PC or GPU box running Ollama replaces private drafting and light code help for a few euros a year in electricity — it doesn't replace frontier reasoning.
A music server (Navidrome) needs 50–100MB of RAM and is worth it the moment you own the files you're streaming.
n8n self-hosted is free with unlimited executions; the real floor is 4GB RAM, not the official 2GB minimum.
A self-hosted Git host (Gitea/Forgejo) runs comfortably under 1GB RAM and is a clean GitHub replacement for personal or small-team repos.
Not every category is worth it — a self-hosted Reddit reader is the one spoke in this series where "just use an ad blocker" beats running your own server for most people.

KEY-STAT: 6 — Big-tech app categories tested on this site's own homelab for this series, each with a real hardware and time cost, not a feature list

The self-host-or-not decision table

Replaces	Self-hosted option	Real hardware cost	Verdict
ChatGPT (private drafting, light code help)	Ollama	Mini PC/GPU box; a few €/year electricity	Worth it for privacy and light tasks — not frontier reasoning
Spotify / Apple Music	Navidrome	50–100MB RAM, negligible CPU	Worth it once you own the files
Zapier / Make	n8n (self-hosted)	300–500MB RAM idle, 4GB realistic floor	Worth it unless you need SSO or Git-based workflow versioning
Official Reddit app	Redlib	Single Docker container	Only if ads genuinely bother you enough to run your own instance
GitHub (repo hosting)	Gitea / Forgejo	Under 1GB RAM	Worth it for personal or small-team repos
GitHub Actions cloud minutes	Self-hosted runner	Any spare homelab box	Worth it for private repos with heavy CI use; needs care on public repos

Running your own AI chatbot

A Raspberry Pi 5 handles 1–3B parameter models (Gemma 3 1B, Llama 3.2 3B) at a usable 8–22 tokens/sec; a 7B model on that hardware drops below 1 token/sec and is unusable. Step up to an N100/N150 mini PC and a 7B model runs CPU-only at 6–9 tokens/sec — readable, not fast. A GPU or Apple Silicon box is the tier that actually feels like ChatGPT, hitting roughly 40 tokens/sec on a 7B model. None of these tiers cost more than a few euros a year to run 24/7; electricity was never the expensive part. The full model-by-model benchmarks are in our self-hosted LLM comparison.

Replacing Spotify or Apple Music with your own library

Navidrome is the lightest of the three music-server options tested — 50–100MB of RAM, negligible CPU, built specifically for music, and it speaks the Subsonic API most mobile apps already expect. Jellyfin covers music too, but as part of a full media server that recommends 4–8GB RAM; if you're already running it for video, its music support is good enough to skip a second service entirely. The full three-way comparison, including which mobile apps talk to which server, is in Navidrome vs. Airsonic vs. Jellyfin.

Tip

This category is the clearest "worth it" on the list — once you own the music files, there's no ongoing licensing dependency left to replace.

Workflow automation without a SaaS subscription

n8n's self-hosted Community Edition is free with unlimited workflow executions and the full integration catalog — the software isn't the limiting factor. What it lacks are Enterprise-only features: SSO/LDAP, separate environments, and Git-based version control for workflows. Expect 300–500MB RAM at idle, spiking to 1–2GB during heavy runs; treat 4GB as the realistic floor for a homelab box, not the official 2GB minimum. Full setup notes and the free-tier math that makes self-hosting worth it are in our self-hosting n8n guide.

A self-hosted Reddit reader and your own Git host

Redlib — the maintained fork of the archived Libreddit — runs as a single Docker container and strips ads, trackers, and JS from Reddit by default. But public Redlib instances get rate-limited and blocked by Reddit regularly, and self-hosting your own copy doesn't make you immune to that, it just makes the problem yours to manage instead of a stranger's. Full details are in our self-hosted Reddit reader writeup.

Git hosting is a cleaner case. Gitea and Forgejo are both Go-based, share a common codebase (Forgejo is a 2022 governance fork of Gitea, now stewarded by Codeberg e.V.), and run comfortably under 1GB of RAM. GitLab CE is a different class of tool entirely — a full DevOps platform with a documented 4GB RAM minimum. Both Gitea Actions and Forgejo Actions use GitHub-Actions-style YAML, and most workflows carry over with little editing. The full comparison, including GitHub migration friction, is in Gitea vs. Forgejo vs. GitLab CE.

Running your own CI/CD instead of paying for cloud runners

A self-hosted GitHub Actions runner is a small agent installed on your own machine that polls GitHub for jobs — about 10 minutes to set up. GitHub-hosted runners for private repos ship with half the CPU and RAM of the public-repo default (2 cores/8GB vs 4 cores/16GB) and a finite free-minutes allowance; a homelab box typically offers more of both, plus persistent build caches that GitHub-hosted runners reset every run.

Warning

GitHub explicitly recommends against self-hosted runners on public repositories — anyone who opens a fork pull request can potentially execute code on your runner. The fix isn't avoiding self-hosting, it's using required-approval, ephemeral runners, or keeping the runner private-repo-only. The full safe-configuration checklist is in our self-hosted GitHub Actions runner guide.

What's NOT worth self-hosting, yet

The Reddit reader is the honest exception on this list. If you just want fewer ads with near-zero maintenance, old.reddit.com plus a decent ad blocker is the more rational choice for most people — self-hosting Redlib only pays off if you're already comfortable managing a Docker container and want to control your own rate-limit fate instead of relying on someone else's public instance.

The bottom line: where to start

If you already have a NAS or a spare mini PC running, start with whichever category you're already paying a subscription for — music or automation are the lowest-maintenance wins. If you just bought a Raspberry Pi and nothing else, the LLM and Git-hosting categories both run comfortably on that hardware alone. If you have zero hardware yet, a single low-power mini PC covers every category above except a GPU-tier LLM — start there before buying anything specialized.

Frequently asked questions

What is the best self-hosted alternative to big tech apps?

It depends on which subscription you're replacing. For music, Navidrome is the lightest and most tested option. For automation, n8n's self-hosted edition removes SaaS execution limits entirely. There's no single "best" — the decision table above routes you to the category that matches what you're actually paying for today.

Is self-hosting worth it in 2026?

For the six categories tested here, mostly yes — music, automation, AI chatbot, Git hosting, and CI/CD all showed a real, low hardware cost against a genuine subscription or limitation they replace. The one exception is a self-hosted Reddit reader, where an ad blocker is the more rational choice for most people.

What should I self-host first?

Whichever category you're already paying a monthly subscription for. Music (Navidrome) and workflow automation (n8n) have the lowest hardware floor and the clearest ongoing savings of the six covered here.

How much hardware do I need to start self-hosting?

A single low-power mini PC (N100/N150 class) covers music, automation, Git hosting, and CI/CD comfortably. Only running a genuinely fast local LLM (7B+ parameters at real-time speed) pushes you toward a GPU or Apple Silicon box.

Is self-hosting cheaper than paying for subscriptions?

For the categories that showed low RAM/CPU needs — music, Git hosting, automation — yes, the electricity cost is a few euros a year against a recurring subscription. The real cost isn't money, it's the maintenance time, which is why the decision table above weighs both.