Canari research

Measurements taken against the live MCP ecosystem. Each one lists its method, its limits, and the scripts or data behind it. Dates are the date of the snapshot, not the date of publication.

How to read these. Every number below is a point-in-time observation with a stated population. Nothing here is a certification, a grade of any server, or a claim that a listed component is safe. Where a figure is a string relation rather than a proof, it says so.

Measurements

ReportHeadlineDate
Server Card adoption
experiments/server_card_adoption
Of 27,064 reverse-DNS servers with a remote endpoint, 95.07% serve from a domain their namespace implies. Probing all 889 hosts of the 1,335 that do not: 2.81% publish a Server Card or AI catalog (control 8.36%, chi2=21.51, p<0.001). Of 54 real cards, 0 reference a schema URL that resolves.2026-11-16
What an awesome-list listing is worth Sampled 294 repositories added to the largest MCP awesome list (96K stars, 5,189 merged "Add" PRs), each merged at least 30 days earlier: median 1 star, 33% at zero, 72% at five or fewer. Rejected entries from the same window are statistically indistinguishable.2026-11-17
Who actually gets traction in this category Of the 15 most relevant MCP security tools: organization-owned repositories have a median of 1,119 stars, individual-owned 82.5, and individual-owned excluding the single success case 0. That one case is a sub-project of a 304,751-star repository by the same author.2026-11-17
Do MCP clients verify the server they connect to?
experiments/client_census
Nine open-source MCP clients and SDKs, 10,158 source files read by hand: 0 pin a server's identity, warn on endpoint change, or verify an attestation. The single existing defence is the official SDKs refusing cross-origin redirects, which 6 of 9 inherit.2026-11-16
A name detector's operating point is a property of its corpus
experiments/corpus_benchmark
Same detector, corpus 200 → 1,617: precision 0.985 → 0.940, false positives 18 → 74, recall unchanged. Self-calibrating the threshold buys the precision back and charges 10.6 points of recall.2026-11-16
Package existence audit
data/package-existence-20261117.csv
Live figure: see the weekly audit (regenerated every Monday; the most recent run is the authoritative number).
First snapshot (2026-11-16): 274 of 22,526 npm/PyPI references (1.22%) pointed at a version that could not be installed — 151 package gone, 123 version not published. The next day’s run gave 275 of 22,657. The registry grows, so always quote the snapshot date with the number.

Recurring: package existence audit — regenerated weekly by a scheduled run; earlier snapshots stay in history/.

Competitor teardowns

SubjectWhat it isReading
mcpcheckupScan and score a single MCP serverteardown · textbook
trustcheckRisk-factor scoring for MCP serversteardown · textbook
MCP QueenContinuous live probing + operational grades + trust receipts for the whole registry; free public API and an MCP endpoint of its ownteardown

Field notes

NoteDate
Honest gap analysis — what Canari does not do2026-11-08
Competitive landscape2026-11-08
Competitor deep research v52026-11-09
End-to-end data flow trace2026-11-09
Audit rounds 13-142026-11-09

Paper snapshots

Local HTML copies of papers this work is positioned against, kept so the numbers in the reports above can be checked without depending on a third-party host.