We froze a SHA-256-anchored census of the public MCP ecosystem on 2026-10-12 (frozen release v1.0.0, Zenodo DOI 10.5281/zenodo.23333514): 42,163 servers, 229,827 tool parameter schemas under the frozen counting rules. Every number below recomputes from the frozen exports — the screening script is CC0, so you can redo the math without trusting us.
(Our live repo now tracks 229,790 tool records with input_schema from 13,119 servers of the 42,163 surveyed, under a stricter in-repo counting rule — different rule, same population. We label both because two counting rules side by side without labels is how bad stats happen.)
The slice that matters — numbers from the frozen snapshot:
- 3,456 servers ship a funds-movement surface. 1,301 of them have write polarity — tools that can move money, not just read balances.
- 2,663 servers combine money and credential handling in one trust boundary. If a prompt injection lands on one of these, the blast radius is "spend" not "read".
- 9,761 servers touch account-identity data in tool parameters.
The uncomfortable part: this is only the public ecosystem. Enterprise-internal MCP deployments — invisible to any census — are where the regulated traffic actually lives.
What we did about it (all CC0 or MIT, take what you need):
- A risk gazetteer schema where every label carries an evidence chain — github.com/DSHCorrectover/mcp-risk-gazetteer
- 13,054 zero-auth funds tools mapped as an armory of concrete evidence (that's the "who can spend without a credential" map) — github.com/DSHCorrectover/mcp-atlas-census
- A public CI ledger where our own engine's blind spots are documented first — 576 verdict-table rows, replayable from public runs, red runs included: github.com/DSHCorrectover/correctover-recovery-admission/blob/main/INTEGRATION-SURFACE.md
Why this is protocol-agnostic, not an MCP-only story: the verdict layer below the census is protocol-agnostic — the same (principal, scope, side_effect_class, freshness) → verdict shape lands for OpenAI function calling, Google ADK, LangGraph, custom runtimes. It's not just a claim: that shape is currently under human review inside google/adk-python (PR #8198) — a live thread you can read, not a marketing line.
And one open invitation: we put our own legacy engine's full assessment (a "verdict dossier" — six sections, replay commands, hashes) in public first: github.com/DSHCorrectover/ecr-arena/blob/main/samples/first-report.md. If you run agents that move money, the question worth asking your vendor isn't "can it block?" — it's "who re-authorizes the retry, and can your auditor replay the answer?"
Census: CC0. Gazetteer: CC0/CC-BY. Checker: MIT. Use, cite, or tear into any of it — a red run you can reproduce beats any README.
Top comments (0)