The Nerq Census · No.1 · MCP ecosystem

The MCP ecosystem is healthier than the circulating numbers suggest

A full census of 28,824 public MCP-server repositories finds ~12% dead, ~19% of the living stale (no commits in 180 days), and ~71% actively maintained. Roughly 1 in 9 repositories with public code is deleted or inaccessible — the first continuously-measured mortality reading for the Model Context Protocol ecosystem.

Published 2026-07-22 · Data: nerq.ai · CC-BY dataset · The Nerq Census series

Of 28,824 public MCP-server repositories, ~12% are dead (deleted or inaccessible), ~19% of the living are stale (no commit in 180 days), and ~71% are actively maintained — “abandonment” and “death” are different measures, and by both the MCP ecosystem is broadly healthy, not collapsing.

The Nerq Census No.1 classifies 28,824 distinct public GitHub repositories linked from indexed MCP servers (99.86% classified). Dead (repository deleted or made private): 11.98% of classified (3,448). Stale (alive, no commit in 180 days): 19.52% of living repositories (4,945). Actively maintained (commit within 180 days): 70.84% (20,390). Dead-or-stale tops out near 29.16%. Two independent methods (commits.atom census + a GitHub REST API random subsample) agreed on 98.8% of overlapping repositories. Data published under CC-BY-4.0. Last analyzed 2026-07-22.

Two different questions, two different numbers

Over the past months a striking figure has circulated about the Model Context Protocol ecosystem — that a majority of MCP servers are “abandoned.” That figure measures staleness: whether a repository has recent commits. It is a real and useful measure. But it answers a different question than the one most people hear in the word “abandoned,” which is whether the project is gone.

Death and staleness are distinct, and they produce very different numbers. So we measured both, defined each precisely, and stated every denominator — across a census an order of magnitude larger than earlier one-off samples.

Death

The repository is deleted or inaccessible — its public URL returns 404/410. (This means deleted or made private; the code may live on privately. We say “deleted/inaccessible,” never assume “gone forever.”)

Staleness (commits_stale_180d)

The repository is present and reachable but has no commit in 180 days. A stale repo is not a dead repo — the code is there, installable, and may be complete rather than abandoned.

What the census found

Of 28,824 distinct public GitHub repositories linked from indexed MCP servers, 99.86% were classified:

Read against a single denominator: counting both the dead and the stale, at most ~29% of classified repositories are dead-or-stale — and only ~12% are actually gone. The large majority, ~71%, show commits within the last six months.

Widely-shared “majority abandoned” figures were built on a different criterion (staleness) and a much smaller sample (on the order of ~1,800 servers) than this census of 28,824. Neither number is wrong for what it measures; they measure different things, on different denominators, at different scales. Stated side by side, the ecosystem looks less like a graveyard and more like a young, working software ecosystem with a normal long tail of finished and paused projects.

Methodology

Two independent measurement paths

Liveness was 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 death signal. To validate that method, a random subsample was independently re-measured against the authoritative GitHub REST API (pushed_at + archived). The two paths agreed on 98.8% of the overlapping repositories (the agreement N grows with each cross-check pass); the residual disagreements are boundary cases within days of the 180-day line.

404 semantics

GitHub 301-redirects renamed or moved repositories, so a 404 means the repository was deleted OR made private — not necessarily that the code is gone. We report “deleted/inaccessible” and never overclaim.

Censoring, disclosed

This is a prevalence census — the state of each repository now — not a lifespan study. Two honest limits: MCP-server “births” are left-censored (our index records when we first saw a server, not when it was created), and the daily-snapshot backbone has two irrecoverable observation gaps (24 and 34 days) that any death inside them carries as an interval, never a point. For those reasons we deliberately publish no median lifespan yet; a second dated sweep (already accruing) is the minimum needed to estimate it.

An archive number that gets more precise over time

This census is one reading from a continuously-measured panel. Every weekly sweep re-probes the small unclassified tail and reclassifies it, so the bounds on every figure tighten week over week. A one-off audit's number is frozen at its error bars forever; a panel's number improves with every sweep. That is the property we are building.

Scale figures in this report are counts from Nerq's own index of the MCP ecosystem as of 2026-07-22 and are reproducible from the published dataset; they are not claims about the total universe of MCP servers that may exist outside any public index. *Archived share is from the GitHub-API cross-check subsample (small N; see the dataset for the per-repo record).

The data is public

The full per-repository classification is published as a versioned, downloadable dataset under CC-BY-4.0 — free to use with attribution to Nerq. One row per repository (repo, source_url, state, last_commit), plus a summary and data dictionary.

Download the dataset (JSONL, CC-BY)

Frequently asked

Part of The Nerq Census — a continuously-measured panel of the AI-agent economy. Each sweep tightens the numbers.