csoai/mcp-remote-census

Remote MCP endpoint census - measured read, 2026-09-25

What remote MCP endpoints listed in public catalogues answered to initialize and tools/list, on 2026-09-25T06:38:51Z to 2026-09-25T06:54:10Z, from one network location. Published by Council of AI (CSOAI). We measure. We do not certify, grade, rank or endorse any server.

Read state: PARTIAL. 4,144 endpoints attempted of 4,376 planned (232 not attempted: robots.txt). The plan is the top 20% of an ordered frame; every count below is over the attempted endpoints and is not a population total of MCP servers.

Record: record.json sha256 c4880fdba78709034466045f6c2ac1eda7c466c96502c75cca68ea34795dce84, signed (Ed25519, did:web:csoai.org#board-attestation-1) at 2026-09-25T07:07:20.977Z, OpenTimestamps state PENDING_CALENDAR_COMMITMENT at publication (see below).

Corrections

2026-09-25: record 0.1.1 supersedes 0.1 (how to read protocolVersion)

record.v0.1.1.json (sha256 fd5c8a7a65f23f7080f2859e386afe474f8a2d1d8316e5be60be4304dff7118a) supersedes record.json (0.1, sha256 c4880fdba78709034466045f6c2ac1eda7c466c96502c75cca68ea34795dce84). The 0.1 record, its signature and its timestamp proof are kept here unchanged. Scope: interpretation of protocolVersion only; no measurement re-run; every count, state, row and published file identical to 0.1. Text further down this card that repeats the 0.1 wording is superseded by this section.

What was wrong

• anomalies[A1].what_we_do_not_say said: "anything about the operator's intent, or whether the version string is correct". It left open whether '2026-07-28' is a real MCP protocol version. It is the current published MCP specification revision, released 2026-07-28, 59 days before this probe. We had not checked; one fetch settles it.
• protocol_version.note (README: 'protocolVersion answered') said: "a server answers with the requested version when it supports it, otherwise with one of its own; this distribution is conditional on the request and is not a census of what servers support". It stated the legacy negotiation rule without saying that the version we requested, 2025-11-25, is itself a legacy revision and that we sent nothing in the current one. The distribution could be read as describing what current servers support; it describes what they negotiated against an older request.

What is right

• 2026-07-28 is a published MCP specification revision; the specification site labelled it '(latest)' when fetched (evidence below).
• It removes the initialize/initialized handshake and sessions: each request carries its protocol version in _meta and, on HTTP, the MCP-Protocol-Version header; servers MUST implement server/discover; an unsupported version is answered with UnsupportedProtocolVersionError (-32022) listing the supported ones. Revisions 2025-11-25 and earlier are 'legacy'.
• Our 0.1 prober requested 2025-11-25 through a legacy initialize only. The protocolVersion distribution (2025-11-25: 1,419, 2025-06-18: 297, 2025-03-26: 116, 2026-07-28: 95, 2024-11-05: 80, 2025-11-05: 1) is therefore servers' negotiation against an older requested version, not their maximum supported version. Under the 2025-11-25 lifecycle a server that supports the requested version MUST answer with it, so the 1,419 endpoints that answered 2025-11-25 include any that also support 2026-07-28; this probe could not see which.
• The 95 endpoints that answered '2026-07-28' (94 getvda.ai hosts, plus oceanalt.com) named it inside a legacy initialize exchange, a handshake that revision retired. Under the 2025-11-25 lifecycle a server answers a version other than the one requested only when it does not support the requested one. We record what they answered and do not say which revision's behaviour they implement.

Primary sources (fetched, hashed, quotes checked against the fetched bytes; evidence/mcp-2026-07-28-primary-sources.json in the repo, sha256 e4a50b37b3f7215b477bee3fdc4a526a4e6baa8985dd18a8a6142607803d8099):

URL	fetched (UTC)	sha256 of bytes	quoted
https://blog.modelcontextprotocol.io/posts/2026-07-28/	2026-09-25T08:04:56Z	9e3c347f359baa66725fffe8e88f348299905d236dea320cd8457e8c4619e520	"pushing the release button on the next version of the MCP specification, 2026-07-28" / "officially retired the initialize / initialized exchange"
https://modelcontextprotocol.io/specification/2026-07-28	2026-09-25T08:04:57Z	e434010502918455ebe12ede0d8719c3e3dec25ed9655d2a7fc0e4944de021c7	"Version 2026-07-28 (latest)"
https://modelcontextprotocol.io/specification/2026-07-28/basic/versioning.md	2026-09-25T08:04:57Z	28c417b3c38345ae8350b9be20ed1b53fcec564b65c13c25d96ab9cfe7e44e9f	"There is no negotiation handshake." / "establish a session with an initialize handshake (2025-11-25 and earlier)"
https://modelcontextprotocol.io/specification/2026-07-28/basic/transports/streamable-http.md	2026-09-25T08:04:57Z	761f1cd04bfdc4f2c16a362502725ea684341cb7ad99607df927017713af19fc	"Every POST request to the MCP endpoint MUST include an MCP-Protocol-Version header."
https://modelcontextprotocol.io/specification/2025-11-25/basic/lifecycle.md	2026-09-25T08:04:58Z	884720004be3ab0e2c65562c33ce8903b30e7c763c4764baed04a2ef8dfd7fa6	"Otherwise, the server MUST respond with another protocol version it supports."

Prober. scripts/census/mcp-remote-probe.py (csoai.census-probe/0.2, commit 61da72861) offers 2026-07-28 first (server/discover with _meta and modern headers), follows -32022 to a listed version it speaks, never falls back past a recognised modern error, otherwise falls back to a legacy initialize requesting 2025-11-25; each row records protocol_versions_offered, protocol_version_requested, protocol_version_negotiated and era separately. It has not been re-run as a census.

Smoke test of the corrected prober (5 endpoints, not a census; these rows are all of it):

endpoint	era	requested	negotiated	server lists	answered to the 0.1 initialize
https://mcp.context7.com/mcp	modern	2026-07-28	2026-07-28	2026-07-28	2025-11-25
https://mcp.svelte.dev/mcp	modern	2026-07-28	2026-07-28	2026-07-28	2025-06-18
https://www.basealpha.fun/mcp	legacy	2025-11-25	2025-03-26	-	2025-03-26
https://oceanalt.com/api/mcp	modern	2026-07-28	2026-07-28	2026-07-28, 2025-06-18	2026-07-28
https://auth-token-mcp-service.getvda.ai/mcp	legacy	2025-11-25	2026-07-28	-	2026-07-28

Verify 0.1.1. record.v0.1.1.signed.json: Ed25519 by did:web:csoai.org#board-attestation-1, signed 2026-09-25T08:05:36.586Z, payload sha256 e72452958a012bc2e9d5fed50505ef090fde0b1a65fd4b7d9eb22ef904aa5d54; verify exactly as for 0.1 below, with the 0.1.1 file names. record.v0.1.1.json.ots: 3 pending calendar attestations at publication, PENDING_CALENDAR_COMMITMENT, not a Bitcoin attestation.

1. The frame (what the catalogues list)

source	read_state	rows read	distinct endpoints	population total
mcp-registry	EXHAUSTED	35,873	21,958	35,873
hf-spaces	EXHAUSTED	10,822	10,822	10,822
a2aregistry	EXHAUSTED	443	440	443
docker-mcp-registry	EXHAUSTED	328	77	328
smithery	PARTIAL	500	0	null

Endpoints read across sources: 33,261 on 25,526 hosts. Combined population total: null, because smithery: PARTIAL. endpoints_read counts what was read. It is not a population total: Smithery was read PARTIAL (264 distinct of a declared 17,186; the anonymous API stops at 5 pages and lists no endpoint URL).

An independent second read of the MCP registry (cited, not merged)

Two readers of the official MCP registry, written separately, ran about an hour apart on the same day. Their row counts are CITED side by side. The data is not merged, and this is not a reconciliation: the registry changed between the reads, so a difference is expected and is not an error in either.

reader	user agent	endpoint	when (UTC)	read state	pages	rows	summary sha256
census frame (this record)	CSOAI-census/0.1 (+https://councilof.ai)	https://registry.modelcontextprotocol.io/v0/servers?limit=100&version=latest	2026-09-25T05:45:50Z to 2026-09-25T05:54:05Z	EXHAUSTED	359	35,873	ce4bb52a70a6077c3f2ab004e04da616578bcad362bf712e86f63c9da48428e2
RAS discovery lane (a different lane, different code)	CSOAI-RAS-Discovery/0.1	https://registry.modelcontextprotocol.io/v0.1/servers?limit=100&version=latest	finished 2026-09-25T04:40:39Z (file mtime)	COMPLETE	359	35,859	8ec4692760de6a30b9687704fba33c36327d70a2258bccb16da33f0b4dd323f2

Time gap: the RAS read finished at about 04:40Z; the frame read ran 05:45:50Z-05:54:05Z, roughly 65-74 minutes later. Ids in both: 35,859; only in the frame read: 14; only in the RAS read: 0; shared ids with a different latest version: 29. Every id in the earlier RAS read (35,859) is also in the later frame read, which has 14 more; 29 shared ids carry a different latest version. That is consistent with publications made between the reads, and the two reads agree to within that churn. 'remote' (RAS) and 'entries with an endpoint' (frame) are defined differently and are not compared. The RAS lane also diffed its read against the estate's canonical public MCP corpus (mcp-corpus-diff-20260925.json, sha256 6c8c41b133320154acd558cf34e1710bdbbb3f90d97b37dfc735f54ddf82e116): 4,364 new, 312 absent from the current registry, 3,509 version changes against 31,807 canonical ids.

2. The plan (which 20% we probed, and why that 20%)

21,880 probeable candidates; top 4,376 planned.

• Tier 1, reach-signalled (1,575): dockerhub_pull_count 16, npm_weekly_downloads 1,276, smithery_use_count 6, pypi_downloads_7d 277. ordered by standing WITHIN the endpoint's own signal (reach_pct); signals are different units and are never added or converted.
• Tier 2, ALPHABETICAL FILL - not a reach signal (2,801): the remaining plan slots filled in MCP registry listing order, then Docker catalogue order. Registry listing order is byte-ascending server name (35,872 of 35,872 adjacent pairs ascending); the fill stopped at com.backgroundcheckquotes/site. io.github.* names: 0 of 2,801 in the fill vs 11,324 of 21,705 registry entries with a probeable endpoint. the fill favours reverse-DNS names early in the byte order; it is a deterministic tie-break, not a sample of the registry.
• GitHub stars were not used: api.github.com serves 60 unauthenticated requests/hour; no key is used by rule.

3. The probe (what the endpoints answered)

state	endpoints	meaning
RESPONDED	2,008	answered initialize with an MCP result
AUTH_REQUIRED	1,428	refused without a credential (401/402/403 or an auth error); no credential was ever sent
NOT_MCP	346	answered, but not with MCP (HTML, 404, non-JSON-RPC)
UNREACHABLE	255	no usable answer: DNS, TLS, connect failure, or HTTP 5xx
TIMEOUT	62	no answer within 10 s connect / 15 s read
SSE_ENDPOINT_ONLY	37	legacy SSE transport: served an endpoint event; no session was opened
MCP_ERROR	8	answered with a JSON-RPC error to initialize

Sent: GET /robots.txt (once per host); POST initialize; POST notifications/initialized; POST tools/list (<= 5 pages); DELETE (only when a session id was issued); GET with Accept: text/event-stream (legacy SSE; endpoint event only). Never sent: tools/call, resources/read, prompts/get, any credential, any payment. One connection and at least 1.0 s between requests per host; at most one retry. 12,739 requests in total. Correction recorded: re-probe after a client fix: an SSE data line longer than one 64 KiB readline was split and dropped, so large tools/list results over SSE read as no_response (93 re-probed; first run kept, not published).

Tool surface (RESPONDED, tools/list complete)

Units: tools listed by one endpoint. Percentiles by nearest rank. A sum of tool counts is not a population figure.

view	n	p25	median	p75	max
per endpoint	1,983	4	7	16	627
per host (best-ranked endpoint)	1,781	4	7	13	627
per endpoint, excluding getvda.ai and a2awire.com	1,777	4	7	14	627

1,701 distinct tool-name sets; the largest identical set covers 92 endpoints.

protocolVersion answered (we requested 2025-11-25)

protocolVersion	endpoints
2025-11-25	1,419
2025-06-18	297
2025-03-26	116
2026-07-28	95
2024-11-05	80
2025-11-05	1

a server answers with the requested version when it supports it, otherwise with one of its own; this distribution is conditional on the request and is not a census of what servers support.

4. Anomalies

• A1 - 92 distinct getvda.ai subdomains answered with the identical 2-tool set ['invoke', 'selftest'] and every getvda.ai endpoint answered protocolVersion '2026-07-28' to a request for '2025-11-25'. They are 94 of the 2008 RESPONDED endpoints and the single largest identical tool-name set. Counts per endpoint therefore over-weight one operator.
• A2 - 109 of 112 a2awire.com endpoints fall into two tool-name sets (59 with 16 tools, 50 with 23 tools). One host is the largest single contributor of RESPONDED endpoints.
• A3 - 231 attempted endpoints are listed at hostnames that do not resolve (EAI_NONAME/EAI_NODATA at the robots.txt fetch); 1 more had a temporary resolver failure and is not counted as dead. 39 of the 231 are our own (see own_entries).

5. Our own dead entries

39 remote endpoints published under our own namespace io.github.CSOAI-ORG in the official MCP registry point at api.meok.ai, which does not resolve in DNS. They were probed like everyone else's and are counted in UNREACHABLE above. This is our defect, not the registry's. Deprecating or deleting any io.github.CSOAI-ORG/* entry requires GitHub authentication as an Owner of the organisation; a domain proof for councilof.ai grants ai.councilof/* only. The reasons, from the registry's own source, are recorded in the supersession record. Record: https://councilof.ai/interop/mcp-registry-2026-09-23/supersession.json (sha256 7802f6641c41ce2deb8b3c16cebfa064cd1dd86638d2567d7d08a3e20ef9fb81). To exclude the measuring estate's own rows: drop rows whose host is api.meok.ai or councilof.ai.

6. What this is not evidence of

• that any server is safe, secure, correct, maintained, or fit for any purpose
• the quality of any server or its tools, or that a tool does what its name says
• endorsement, ranking, grading, or approval of any server or operator
• anything about the ~80% of the frame outside the top-20% plan, or the 232 planned endpoints not attempted
• a population total of MCP servers or endpoints: the frame is PARTIAL (Smithery) and the probe is PARTIAL
• reachability at any other moment or from any other network location
• what AUTH_REQUIRED servers expose behind their credential

Files

file	sha256
record.json	c4880fdba78709034466045f6c2ac1eda7c466c96502c75cca68ea34795dce84
probe/results.public.jsonl.gz	f29c89c910f1677ab570de701494415806a9796770323e5c50a79fc2664b2bb3
probe/not_attempted.jsonl.gz	663cc83fa46fbe103f69a540c6ab1ddcc43374328e0b5e735c1f7b68b14a6be1
frame/summary.json	ce4bb52a70a6077c3f2ab004e04da616578bcad362bf712e86f63c9da48428e2
frame/plan-top20.json	493c81c10a2b91c0707e2df3e6e6c07f614898d881282ae18766b703c8d53565
probe/summary.json	991ef00bf40c6c56960ff08110fdf3ebceb366f5d1f8b15aadf3f85cb8ef9ae1

record.signed.json (signature), record.json.ots (OpenTimestamps proof), record.ots.json (its stated state). Per-endpoint rows carry: rank, ranked_by, endpoint, host, state, reason, http_status, transport, protocol_version, tools_list_status, tools_complete, n_tools, tool_names_sha256, started, finished. They carry no tool names, tool descriptions or serverInfo; tool_names_sha256 is sha256 of the sorted tool names joined by \n, UTF-8 (as the probe computes it). Rows are derived from public catalogue listings.

How to verify

Signature. In record.signed.json: serialise payload as JSON with keys sorted recursively, no whitespace, UTF-8. Its sha256 must equal signature.payload_sha256. payload.artifact.sha256 must equal sha256 of record.json. Verify signature.sig_ed25519 (hex) over those bytes with the Ed25519 key #board-attestation-1 (publicKeyJwk.x, base64url) in https://csoai.org/.well-known/did.json. Change one byte of the payload and it must fail.

import json, hashlib, base64, urllib.request
from cryptography.hazmat.primitives.asymmetric import ed25519
s = json.load(open("record.signed.json"))
c = json.dumps(s["payload"], sort_keys=True, separators=(",", ":"), ensure_ascii=False).encode()
assert hashlib.sha256(c).hexdigest() == s["signature"]["payload_sha256"]
assert hashlib.sha256(open("record.json", "rb").read()).hexdigest() == s["payload"]["artifact"]["sha256"]
did = json.load(urllib.request.urlopen("https://csoai.org/.well-known/did.json"))
x = [m for m in did["verificationMethod"] if m["id"].endswith("#board-attestation-1")][0]["publicKeyJwk"]["x"]
ed25519.Ed25519PublicKey.from_public_bytes(base64.urlsafe_b64decode(x + "==")).verify(bytes.fromhex(s["signature"]["sig_ed25519"]), c)

Timestamp. pip install opentimestamps-client, then ots upgrade record.json.ots and ots verify record.json.ots (verification needs a Bitcoin node or a block explorer). At publication the proof held 3 pending calendar attestations from https://alice.btc.calendar.opentimestamps.org, https://bob.btc.calendar.opentimestamps.org, https://finney.calendar.eternitywall.com: a pending calendar commitment, not a Bitcoin attestation. It proves nothing about time until it is upgraded and verified against the chain.

Licence

Data: CC-BY-4.0. Cite as: Council of AI (CSOAI), Remote MCP endpoint census, measured read 2026-09-25, csoai/mcp-remote-census.

Sibling censuses

Two further measured reads from the same 2026-09-25 census frame, each with its own signed record and OpenTimestamps proof (pending calendar commitment at publication). They are separate populations and are never added to the figures above.

dataset	what it measures	read state	record.json sha256
csoai/hf-mcp-spaces-census	the 10,822 Hugging Face Spaces tagged mcp-server: Hub API stage for all; initialize + tools/list only for the 2,658 RUNNING Gradio Spaces (sleeping, paused and error Spaces are never requested)	EXHAUSTED	53022d26ca321b291ca5578b3b318ef645ecf1e4343b13320ebc91642abb906e
csoai/a2a-card-census	the 443 a2aregistry listings: agent card served, and card signatures checked under the key the card points to	EXHAUSTED	b290b53d05912d8b3a5913c7e73e693f40cc71be8e0432c4626643c6d662176c

This remote-endpoint census refuses every *.hf.space host; the Spaces census exists so those endpoints are measured without waking a sleeping Space.

Sibling census added 2026-09-25: contract parity

dataset	what it measures	read state	record.json sha256
csoai/mcp-contract-parity	whether one remote MCP service's public surfaces (registry entry, /.well-known/mcp.json, server card, agent card, x402 manifest, live discovery answers) agree on version, tools, auth, payment and protocol; plus a HOLD list of servers whose registry version changed, re-measured	EXHAUSTED (5828 of 5828)	45e3fd63fc98ad251a4f9fe5fdb705ed7fbc28321d5c205d10114a21730cc323

A separate population built on this census's RESPONDED endpoints; never added to the figures above.

Timestamp status

Checked 2026-09-25T13:46Z by ots upgrade on a copy of each published pending proof. A pending .ots is never rewritten: an upgraded proof is published as a NEW file <name>.bitcoin.ots beside it. BITCOIN_ATTESTED means the proof carries a BitcoinBlockHeaderAttestation whose commitment (byte-reversed) equals the merkle root of the block header at that height as served by two independent public sources (blockstream.info and mempool.space; each raw header also double-SHA256s to the block hash it is served under). No local Bitcoin node was used. Details: ots-bitcoin-check-2026-09-25.json.

• record.json (sha256 c4880fdba78709034466045f6c2ac1eda7c466c96502c75cca68ea34795dce84): BITCOIN_ATTESTED at block 968519 (checked against 2 header sources: blockstream.info, mempool.space). Upgraded proof: record.json.bitcoin.ots (sha256 7ba5fd4a0b39cd188b4b0df0eaa4426046790303391f4080ae943c22e15d1170); the published pending proof record.json.ots is unchanged.
• record.v0.1.1.json (sha256 fd5c8a7a65f23f7080f2859e386afe474f8a2d1d8316e5be60be4304dff7118a): BITCOIN_ATTESTED at block 968527 (also blocks 968533, 968537, via other calendars) (checked against 2 header sources: blockstream.info, mempool.space). Upgraded proof: record.v0.1.1.json.bitcoin.ots (sha256 b95f283e37d1e7624302b01ed823c96576eec0320eef286c9faa5808061a5a55); the published pending proof record.v0.1.1.json.ots is unchanged.
• Reconciliation: an earlier same-day upgrade of the same pending proofs is published under ots-upgraded/2026-09-25/ (checked against one header source). It commits to the same file digests; calendars can return different block paths, so the heights may differ. Either proof suffices; neither replaces the pending original.