CORRECTOVER
runtime verification · evidence infrastructure
Xiamen, CN · est. 2026
Milestone A · A frozen, reproducible census

What MCP servers actually expose — measured at runtime, not declared in metadata.

On October 12, 2026 we completed a full runtime survey of the public MCP registry: every reachable server was handshaked, every open tool surface enumerated, every parameter schema kept as evidence. This report is the graded summary of that census.

26,823
servers probed
13,089
open & reachable
229,539
live tools measured
25,143 10.95%
high-consequence tools

§1Why a census had to be run by hand

Registry metadata says where a server lives. It never says what the server does.

Every published MCP package and registry entry lists package names and remote endpoints. None of them — at any layer — carries tool definitions. Published npm packages don't retain a static manifest either. The only source of truth for a server's tool surface is a live MCP handshake followed by tools/list.

So we connected to all 26,823 candidate endpoints, one at a time, and kept what came back. The two data files behind every number on this page carry SHA-256 manifests you can check yourself, before you pay us anything.

Scope is stated exactly: the public registry snapshot of 2026-10-12. Not a claim of total-ecosystem coverage. Private and unlisted servers are, by definition, not in the frame.

§2What the report gives you

Analysis is our work product; the snapshot is openly licensed. You can reproduce everything before the invoice clears.

Ecosystem risk ranking.

Every open server graded A/B/C against five parameter-risk classes; the top 100 by exposure listed with evidence line references back into the frozen dataset.

Five risk classes, counted.

SSRF · command injection · credential exfiltration · tool poisoning · dependency rug-pull — each high-consequence tool mapped to a class, each count recomputable from the raw lines.

Your surface, against the field.

Hand us your server list and we grade each one on the same frozen baseline: exposure percentile, named findings, and where you sit versus the ecosystem.

Method on the record.

A methodology appendix states classification rules and thresholds plainly. If you disagree with a grade, you can show exactly where and why — that's the point.

§3Rate card

Three ways to buy. Delivery is calendar-days, not quarters.

Tier
What's delivered
Price & terms
Standardone-time
Full ecosystem ranking (top-100 + A/B/C grades), five risk-class breakdowns, methodology appendix, provenance page. PDF + CSV. 3 business days.
rates on requestsingle reportorder →
Customone-time · most picked
Everything in Standard, plus: your server list graded item-by-item, yours-vs-ecosystem percentile per server, named findings with evidence line refs, and a 30-minute findings walkthrough. 5 business days.
rates on requestsingle reportorder →
Monitorsubscription
Monthly re-scan delta against the frozen baseline, new high-consequence tool alerts, continuous tracking of your watchlist, quarterly methodology review.
rates on requestper month · cancel anytimestart →
Provenance
servers.jsonl.gz  ·  8,576,997 B  ·  42,163 rows  ·  sha256 3071866184521fb36aeac6ac715687ed4bf99b9b034646da61176f54a3dfa9d8
server_tools.jsonl.gz  ·  90,047,371 B  ·  229,827 rows  ·  sha256 2350dc773dc43fd6ed0b012588f3ff295fb358106a0291a7dc59c57427f5780f
frozen 2026-10-12  ·  DOI 10.5281/zenodo.23333513  ·  data license CC0 1.0 (repo) / ODC-BY 1.0 (DOI record)
ranking & grading — analysis © 2026 Correctover

§4Order form

We reply within 24 hours with the sample report and an invoice link. Purchase orders and net-30 accepted.

Request
form no. CS-A/2026-10

By submitting you agree to be contacted about this order. No telemetry, no user-data resale. "MCP" refers to the Model Context Protocol ecosystem; census scope is the public registry snapshot of 2026-10-12. IETF drafts referenced on other pages of this site are individual submissions, not RFCs or IETF endorsement.