The State of Self-Hosted Software 2026

The headline

32% of scored self-hosted projects are slowing or at risk — 359 of 1126. That is the figure worth carrying away: pick a project at random from a directory and there is roughly a one-in-three chance its development has materially decelerated or stopped.

It is not all decay, and the ecosystem's reputation for abandonment is unfair. 67% shipped a tagged release within the last ninety days, and only 8% went the whole quarter without a single commit. Self-hosted software is not dying — it is bimodal. A large active core ships constantly, a long tail sits untouched, and the median score of 63 falls in the valley between them.

Which is precisely why an average is useless here and a per-project score is not. The catalogue does not have a health problem; it has a selection problem. The good options and the dead ones sit side by side in every category, indistinguishable until you look at the commit log — and star counts actively mislead, because they only ever go up.

The measurement gap nobody talks about

226 of 1352 projects (17%) cannot be scored at all, because they are developed outside GitHub — on Codeberg, on self-hosted Forgejo or GitLab, or on infrastructure the project runs itself.

There is something quietly appropriate about that: the self-hosting community hosts its own forges, and in doing so becomes invisible to the tooling that measures open-source health. Every ecosystem dashboard you have ever read has this blind spot, and almost none of them say so. We report those projects as unrated rather than guessing, because scoring them from missing data produces exactly one outcome — healthy projects labelled dead.

Method, in one paragraph

Every project is scored 0–100 from public repository metadata: commit volume over 11 complete months (40 points, percentile-ranked against the whole catalogue), days since the last tagged release (25, absolute), community size (20, percentile-ranked) and quarter-on-quarter momentum (15, absolute). Percentile ranking on the two continuous components stops the scale saturating and makes the score impossible to inflate with trivial commits. Archived and unmeasurable projects are excluded from the ranking base. Full method, including its known limitations.

Using these figures

Every number here is recomputed from source on each rebuild, so this page does not go stale the way an annual PDF does. You are welcome to quote, chart or reproduce any of it — a link back to this page is all we ask. The underlying software metadata comes from the community-maintained awesome-selfhosted dataset under CC-BY-SA 3.0 (attribution); the scores and analysis are ours.

Last recomputed 2026-10-01 from source data dated 2026-10-01.