How many listed agent extensions still have a working source repository?

A recurring claim in 2026 is that roughly half of published MCP servers are “dead”. We measured it against our own corpus rather than repeat it.

n = 9,183 · measured 2026-10-19 · recomputed on every page load

Result

Observation	Count	Share
Commits observed in the last 90 days	6,881	74.9%
No commits in over 90 days	822	9%
Source URL did not resolve	1,480	16.1%
Neither recently updated nor resolving	2,302	25.1%

What this does and does not say

It says 16.1% of sampled listings point at a source URL that did not resolve when checked, and a further 9% point at repositories with no commits in over 90 days. It does not say those projects are abandoned, unmaintained or unsafe. A URL that does not resolve may have been deleted, renamed, or made private. Stable software may need no commits. We report the observation; the inference is yours.

Method

Sample drawn at random from artifacts publishing a source repository URL. Random selection matters: listings are not evenly distributed across the alphabet, and an alphabetically ordered scan of this same corpus reported a non-resolving rate of 72%, roughly triple the random-sample figure. That is a sampling artifact. We mention it because it is an easy way to publish a wrong number, and because we made exactly that mistake first.
Queried through the GitHub API with authentication where available. An unauthenticated request returns 404 for private repositories as well as deleted ones, which would report live projects as gone.
Renames are followed rather than counted as failures. Rate-limited checks are skipped, never recorded as failures.
Threshold for “no recent commits” is 90 days since the last push to the default branch.

Independent replication

An earlier ad-hoc audit of 120 repositories drawn only from the official MCP Registry, run through entirely separate code, found 62.0% with recent commits and 37.0% neither recently updated nor resolving. This pipeline, on a larger and differently composed sample, finds 74.9% and 25.1%. In that earlier audit every non-resolving repository was re-checked authenticated and following redirects; 36 of 36 still did not resolve.

Full method and known limitations: methodology. Believe a specific record is wrong? Request a correction.