The Nerq Census · No.2 · MCP ecosystem
The MCP Deadpool: 21.3% of MCP-server repositories are deleted or abandoned
Across 28,824 public MCP-server repositories — the same population as Census No.1, read in four dated verdicts — 12.0% are deleted (404/410) and 9.3% are abandoned (no commit in over a year). Counting repositories dormant 180–365 days, 29.1% are dead or declining; 70.7% are actively maintained.
Published 2026-07-27 · Data as-of 2026-07-22 · CC-BY dataset · The Nerq Census series
Of 28,824 public MCP-server repositories, 21.3% are deleted or abandoned (12.0% deleted / 404, 9.3% with no commit in over a year), and 29.1% are dead or declining once repositories dormant 180–365 days are included; 70.7% are actively maintained. This is the same 28,824-repository population as Nerq Census No.1, read in finer verdicts — the two independent cuts give the same answer.
The Nerq Census No.2 (“The MCP Deadpool”) reclassifies the 28,824-repository MCP-server population of Census No.1 into four dated verdicts. Deleted (404/410): 3,448 (12.0%). Abandoned (present, no commit >365d): 2,686 (9.3%). Dormant (present, no commit 180–365d): 2,259 (7.8%). Alive (commit within 180d): 20,390 (70.7%). Deleted-or-abandoned = 21.3%; dead-or-declining = 29.1%. 69% of verdicts are exactly dated; 31% are interval-censored (last sign of life falls in a crawl-blind observation gap) and flagged in the schema. Data CC-BY-4.0, as-of 2026-07-22.
Bridge from Census No.1 — same population, finer verdicts
No.2 does not revise No.1. It is the identical 28,824-repository population, reclassified from No.1's coarse dead/stale/active into four dated verdicts. The arithmetic reconciles exactly:
| Reconciliation | No.2 arithmetic | No.1 reported |
|---|---|---|
| Living repositories | 28,824 − 3,448 deleted = 25,376 | — |
| Stale of living (dormant + abandoned) | 4,945 / 25,376 = 19.5% | ~19% |
| Alive share | 70.7% | ~71% |
| Deleted share | 12.0% | ~12% |
Two independent cuts of the same census give the same answer. No.2's contribution is the finer verdict split — separating repositories dormant 180–365 days from those abandoned beyond a year — and dating each transition.
The four verdicts
| Verdict | Definition | Count | Share of 28,824 |
|---|---|---|---|
| Deleted | repository returns 404/410 (deleted or made private) | 3,448 | 12.0% |
| Abandoned | present, no commit in >365 days | 2,686 | 9.3% |
| Dormant | present, no commit in 180–365 days | 2,259 | 7.8% |
| Alive | commit within 180 days | 20,390 | 70.7% |
| Unknown | could not be classified | 41 | <1% |
Headline denominator: deleted-or-abandoned = 21.3% (6,134 of 28,824). Adding the dormant band (180–365d), dead-or-declining = 29.1%. The dormant definition is written out precisely so the objection is pre-empted: a repository quiet for 200 days is counted as dormant, not abandoned.
Interval-censoring — 69% exactly dated, 31% censored, all disclosed
69% of verdicts are exactly dated from the repository's last-commit timestamp.
31% (8,955 repositories) are interval-censored: the last sign of life
falls inside an observation gap when our crawl was blind, so the transition is dated to a window, not a day.
This is disclosed per-row in the CC-BY schema (a censoring column), not smoothed
over.
Interval-censoring is standard in survival analysis: it widens the confidence interval without invalidating the point estimate. The two blind windows are named openly:
| Observation gap | Repositories censored | Cause |
|---|---|---|
| 2026-05-02 → 2026-05-27 | 1,644 | crawl blind window |
| 2026-06-16 → 2026-07-21 | 7,311 | a silent timer death on the collection host — the same gap that interrupted our daily_snapshots drift backbone, since repaired |
Separately, the 3,448 deleted repositories have no recoverable death date at all — a repository that is gone leaves no timestamp behind. We state that plainly rather than impute a date. An explained gap beats an unexplained one; naming the cause is part of the measurement.
Population purity — the mortality is not a tutorial-fork artifact
MCP tutorials and templates are forked widely, which could inflate a mortality count with near-duplicate repositories. We tested it. Removing 1,753 tutorial/template-fork repositories (clustered by near-identical names; 77% of them alive — above the population average) moves dead-or-declining from 29.1% (raw) to 29.5% (purified) — slightly higher, not lower.
Both numbers are published because the purified figure argues against our own headline and
still supports it: the forks skew alive, so removing them cannot be what makes the ecosystem look worse. A
forthcoming refinement will identify forks by identical manifest/config content hashes rather than name
similarity, and every repository will carry a population_class label in the registry.
Typosquatting — a null result, published as a null result
We looked for adversarial typosquatting and did not find it at this layer.
867 name-collision clusters exist (repositories whose normalized names coincide), but they are
dominated by tutorial and template forks — everyone names their weather example
mcp-weather-server — not by adversarial near-names of popular servers. The correct signal is not
name similarity but actual installation: for a given name-twin, how many real configurations
reference it. Zero references is noise; non-zero is a finding. That measurement depends on the configuration
graph and is a later report. We publish this null result rather than dress it up.
Methodology
Verdicts were read from each repository's public commits.atom feed (no API key, not subject to
API rate limits): the most recent entry's timestamp is the last commit; a 404/410 is a deletion signal.
Repositories are the same 28,824-repository population published in Census
No.1 under CC-BY. The verdict thresholds (180-day dormancy, 365-day abandonment) are stated above; the
per-row data — verdict, last commit, and censoring flag — is downloadable under
CC-BY-4.0.
Questions & answers
How many MCP servers are dead or abandoned?
Of 28,824 public MCP-server repositories, 21.3% are deleted or abandoned — 3,448 deleted/inaccessible (404/410) plus 2,686 present but with no commit in over a year. Including repositories dormant 180–365 days, 29.1% are dead or declining.
Does No.2 revise No.1's numbers?
No. It is the same 28,824-repository population in finer verdicts. No.1 reported ~12% deleted and ~19% of living repositories stale; No.2 splits “stale” into dormant (180–365d) and abandoned (>365d) and dates each. The arithmetic reconciles exactly: 4,945 stale of 25,376 living = 19.5%, matching No.1's ~19%. Two independent cuts, same answer.
What is the difference between deleted, abandoned, and dormant?
Deleted = the repository returns 404/410 (deleted or made private). Abandoned = present and reachable but no commit in more than 365 days. Dormant = present, no commit in 180–365 days. They are distinct, dated measures on explicit denominators; none is inferred.
How reliable are the dates?
69% of verdicts are exactly dated from the repository's last-commit timestamp. 31% are interval-censored: the last sign of life falls inside an observation gap when the crawl was blind, so the transition is dated to a window, not a day. Interval-censoring is standard in survival analysis; it widens the confidence interval without invalidating the estimate, and it is disclosed per-row in the CC-BY schema.
Are these numbers inflated by tutorial forks?
No — the opposite. Removing 1,753 near-identical tutorial/template-fork repositories (which are 77% alive, above the population average) moves dead-or-declining from 29.1% to 29.5% — slightly higher, not lower. Both numbers are published; the mortality is not a fork-inflation artifact.
Did you find MCP typosquatting?
Not at this layer. 867 name-collision clusters exist, but they are dominated by tutorial/template forks, not adversarial typosquats. The right signal is actual installation — how many configurations reference the twin — which is a later report; we publish this null result as a null result.
Look up a specific server or your whole catalog
Every deleted, abandoned, or dormant repository has its own dated page — e.g.
picahq/mcp (deleted),
redis/mcp-redis-cloud (abandoned),
chroma-core/chroma-mcp (dormant) —
at /census/server/{owner}/{repo}/. To check an entire catalog at once, paste it into
the exposure surface (no login; only anonymous counts are stored).
Download the CC-BY dataset Check your catalog ← The Nerq Census series