Ashsinha1/mcp-registry-census

MCP Registry Census

A longitudinal census of the public Model Context Protocol registry.

A single snapshot tells you what the registry contains. A series tells you how it is changing, and as far as I can tell nothing else tracks that. This dataset is the series, updated daily.

Files

File	Contents
census-history.csv	One row per census. 13 measurements per run, plus the date.
registry-snapshot-latest.jsonl	Most recent snapshot, distinct servers only (isLatest).
registry-snapshot.jsonl	Same snapshot, every version record.

The one thing to know before using this

The registry endpoint returns one record per published version, not per server. Filter on _isLatest to get distinct servers. Failing to do so changes conclusions by roughly 20x — measured naively the registry looks 74.6% duplicated by description, and 3.6% once deduplicated. The mean is 3.3 versions per server; the maximum observed is 1,175 versions of a single server.

registry-snapshot-latest.jsonl is already filtered. registry-snapshot.jsonl is not.

Columns in census-history.csv

Column	Meaning
version_records	Total records returned by the endpoint
distinct_servers	Records with isLatest set
publishers	Distinct namespaces (the segment before the first /)
single_server_publishers	Publishers with exactly one server
top10_share_pct	Share of servers held by the ten largest publishers
deprecated	Servers marked deprecated but still listed
no_transport	Servers declaring neither remotes nor packages
no_repository	Servers with no repository link
template_placeholder	Servers whose description is still the unedited template
max_versions_one_server	Highest version count for any single server
dup_desc_naive_pct	Duplicate-description rate over version records
dup_desc_correct_pct	Duplicate-description rate over distinct servers

Baseline

First census, 28 August 2026:

• 84,241 version records → 25,423 distinct servers
• 15,709 publishers, of which 14,063 have registered exactly one server
• Ten largest publishers hold 15.5% of the registry
• 394 servers declare no way to reach them — no endpoint, no package
• 268 deprecated servers still listed, 8 still carrying the template description

The registry grows fast. Between an earlier measurement the same day and this baseline, distinct servers rose from 25,125 to 25,423.

Caveats

Counts describe what records declare. No endpoint is contacted, so no_transport is a lower bound on unusable entries — a server declaring a transport may still be unreachable.

The series begins 28 August 2026. Anything before that is not recoverable; the endpoint serves current state only.

Reproducing

git clone https://github.com/ashishsinha1602/dataset-integrity-audit
cd dataset-integrity-audit
python census_track.py          # fetch, measure, append, report drift
python census_track.py --history

Pure standard library, no dependencies.

Related

• Census write-up and discussion: modelcontextprotocol/registry#1580
• The unreachable-servers finding: modelcontextprotocol/registry#1579
• Code: github.com/ashishsinha1602/dataset-integrity-audit

Citation

@misc{sinha2026mcpcensus,
  author = {Sinha, Ashish},
  title  = {MCP Registry Census},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Ashsinha1/mcp-registry-census}
}

MIT licensed.