Repository liveness behind indexed MCP servers — a dated census.
As of 2026-07-22, over n=28,824 distinct public GitHub repositories linked from indexed Model Context Protocol (MCP) servers. One row per observation; each row carries its own last-commit date and HTTP status, so any single classification is re-checkable against GitHub directly. Nothing here is extrapolated beyond this dated sample.
the counts (as of 2026-07-22, n=28,824)
| state | rule | repos | share |
|---|---|---|---|
| alive | last commit ≤ 180 days | 20,390 | 70.7% |
| dormant | last commit 180–365 days | 2,259 | 7.8% |
| abandoned | no commit in > 365 days | 2,686 | 9.3% |
| deleted | an authoritative HTTP 404 | 3,448 | 12.0% |
| unknown | not resolvable at read time | 41 | 0.1% |
| not maintained | abandoned + deleted | 6,134 | 21.3% |
Not maintained — abandoned or deleted — is 6,134 of the 28,824 (21.3%) as of 2026-07-22: no commit in over a year, or gone entirely. Dormant (2,259, 7.8%) is reported separately on purpose — a repository quiet for 180–365 days may be a finished, stable library, not a dead one. If you also count dormant as gone, the combined figure is 8,393 (29.1%); the defensible ‘not maintained’ number is 21.3%. The remaining 70.7% carried a commit within 180 days.
Census date vs. today. These are the dated counts of 2026-07-22. The live register re-reads continuously, so an individual repository's state may have moved since — check the register for its current read. A single-day census dates a state, not a transition.
distributability — how the ecosystem ships
Separately, 62.1% of the 28,824 were not distributable — no importable manifest, or a name-collision cluster where the identifier doesn't resolve to one artifact. This is a property of how these repos are packaged; it isn't a measure of whether they work.
interval censoring (what a single-day census can and cannot say)
A one-day census dates a state, not a transition. When a repository is read as deleted on the census day, its exact death date is bounded, not pinned — it died somewhere between its last successful read and this one. We assert only the dated reads we actually took; a gap between two observations stays a gap, and a change inside an unobserved span is dated to that span, never to a single guessed day. Absence of a record is absence of a record, nothing more.
Observation provenance varies per row. The census date is an as-of stamp, not a claim that every repository was freshly read that single day. Of the population, 57.0% were cleanly observed, 31.1% carry a transition that fell inside an earlier observation gap (2026-06-16 to 07-21) and are interval-censored, and 12.0% were read deleted (death bounded, not pinned). Any later re-census compares like-for-like on the cleanly-observed subset first, and reports the full-population delta only as an upper bound.
Among the cleanly-observed subset, not-maintained is 16.4% (abandonment-only — deletions are inherently death-date-censored and sit outside this subset). The headline 21.3% is not provenance-inflated: its deletion half (12.0%) is measured by a binary HTTP 404 that does not depend on read freshness, and the remainder is dated abandonment.
methodology
- Population = the 28,824 public GitHub repositories linked from the indexed MCP-server census, read first-hand.
- How read: each repository's own public
commits.atomfeed and its HTTP status — no API key required and no private data. Every state is our own dated observation. - Thresholds: HTTP 404 → deleted; otherwise the age of the last commit — ≤180 days alive, 180–365 days dormant, >365 days abandoned. Dead = not alive (dormant, abandoned, or deleted).
- not-checked = a 403 / 429 / timeout / network error at read time, recorded as unknown and never as a death; an outage is not a death.
the live register and the dataset
Each not-maintained repository is a dated leaf in the continuously-updated register: browse the MCP register. The full census dump — one row per observation, with the schema and thresholds — is archived open-access on Zenodo under a permanent concept-DOI (https://doi.org/10.5281/zenodo.21869879) and mirrored on Hugging Face; both are CC-BY 4.0. Cite as: lastseen.dev (2026). Nerq MCP-server repository mortality. Zenodo. https://doi.org/10.5281/zenodo.21869879.
© 2026 lastseen.dev. Every state is lastseen.dev's own first-hand dated observation of a public GitHub record, licensed CC-BY 4.0 — free to use, share, adapt, and use commercially, with attribution to lastseen.dev (link back). These are dated factual observations, without warranty — not a composite score, not a prediction, and not advice about any named party.