Skip to content

Latest commit

 

History

History
423 lines (370 loc) · 30.1 KB

FIRST-WEEK.md

File metadata and controls

423 lines (370 loc) · 30.1 KB

What MCP servers told their agents: the first week of a public log

Data: heldfast-feed at the 2026-09-29 checkpoint [D1]. Code: heldfast at commit b9f7ec8 [C1]. Written 2026-10-01. Each number below was computed by research/report/numbers.py and carries the claim ID that defines it; the register is the appendix.

Summary

  • The log watched 29,458 MCP servers from the official registry: 9,712 npm packages and 19,746 hosted endpoints, so 67.0% of the watchlist was hosted (watchlist of 2026-09-28).
  • Between 2026-09-23 and 2026-09-29 it observed 12,775 changes to tool definitions. With 246 events carried in from an earlier study, the record held 13,021.
  • Three operators accounted for 9,788 of the 12,642 hosted events (77.4%; 75.2% of all 13,021).
  • 66 events were graded review. 58 of them introduced a price, on 41 servers; 8 introduced something else and are listed for a person to read.
  • Of 19,764 hosted servers with a state file, 13,255 (67.1%) had been read and were current. 4,165 (21.1%) had never been read because they answered HTTP 401, 402 or 403; a 403 can be a block rather than a login.

Why hosted servers matter

An MCP client learns what a server's tools do by sending tools/list; each definition carries a name, a description and an input schema, and the model reads the description [4]. Tool poisoning puts instructions there [1], and a server can change its tools after a client has trusted them [2]. A study of open-source MCP servers has measured how often such patterns occur [3].

A package scanner, or an advisory database such as OSV [9], starts from a name and a version. A hosted server has no release to download: it is a URL, and what it returns can change between two requests. On the watchlist of 2026-09-28, 19,746 of the 29,458 servers were hosted. What the log records is the tool list each one returned to an anonymous client, each time it read one. The idea is Certificate Transparency's [5]: after certificates for sites like google.com were issued to people who were not Google [6], certificates had to be logged in public. lastseen.dev keeps a free register of dated observations of MCP servers, including changes to their declared tools, sealed into a daily SHA-256 hash chain [10].

Data and method

Snapshot. The figures come from one commit of rufat325/heldfast-feed, 08a6e310855f6facb9d3bf6e4203db59733bd1ca, the commit the 2026-09-29 checkpoint names [D1]. Its manifest digest is c16aa7cc35e5e04fd65fe2a86d664cb22f102704e0529b0a4a17c3e5e4693e00, over 171,308 files. We rebuilt the manifest from the commit and the digest matched. The OpenTimestamps proof [7] was parsed and commits to the checkpoint file, but was not checked against a Bitcoin node; the Sigstore signature [8] was not checked.

What is read, and how often. Each hosted server is read daily as an anonymous client, and at four-hour intervals while it has changed in the last three days; npm servers are started in a container with no network, the most-downloaded daily and the rest weekly (docs/TRANSPARENCY.md, "What it cannot do"; the watchlist's rule).

What counts as a change. An event is recorded when a server's tool list differs from the last one held: a tool added, removed, or changed in any field the lock digest covers. The collector sorts each event into one of three kinds (research/feed/operators.py): 149 were reorder-only (equal once lists are sorted), 3,150 numbers-only (equal once digits are masked) and 9,722 substantive. Recomputing the kinds from both catalogues of each event gave the same counts as the published stats.json.

How grades are decided. An event is review when a changed tool introduced a signal that the earlier text did not carry: an instruction to conceal, override or exfiltrate, a credential path, a hidden character, a look-alike letter, or a price (heldfast.driftgrade). Otherwise it is quiet. The first review event for a price was observed on 2026-09-27.

What is sealed. Checkpoints began on 27 September 2026. The snapshot's tree holds the checkpoints of 2026-09-27 and 2026-09-28; its own checkpoint was written after it. Records before the first checkpoint were collected the same way but are not anchored.

Results

Volume. The log observed 12,775 changes between 2026-09-23T04:23Z and 2026-09-29T07:38Z: 12,642 on hosted servers and 133 on npm servers. Hosted changes per UTC day were 1,267, 6,028, 257, 2,041, 2,214, 468 and 367 (the last day partial). Without the three largest operators, all observed changes per day were 182, 521, 262, 562, 599, 491 and 370.

Three operators. The record groups hosted servers by registrable domain, which is approximate: it can merge customers of one host and split an operator with several domains. Across 1,260 operators, three stood out. pipeworx.io had 7,023 hosted events on 1,522 servers (55.6% of hosted events): 2,975 were numbers-only and 4,048 substantive, and 7,022 of them changed a description. nolimit-observatory.workers.dev had 2,341 events on 2,341 servers, each one a whole-catalogue change of inputSchema alone. a2awire.com had 424 events on 106 servers; outputSchema changed in 424, inputSchema in 106 and a description in 47.

A catalogue counter that ticks, or a schema regenerated in bulk, is a change the log has to record. It is not misbehaviour, and nothing here says these operators did anything wrong. Corrections are welcome as issues on rufat325/heldfast.

Outside those three. 2,854 hosted events came from 1,288 servers. Across all 5,257 hosted servers with a change, the median was 1 change, and 3,107 servers changed exactly once.

Grades. 12,955 events were quiet and 66 review. 58 review events introduced a price, on 41 servers. Of the rest, by kind of signal (an event can carry several): exfiltration 2, credential path 2, mandated side effect 2, critical word 2, concealment 1, override 1. Those 8 events are listed in research/report/review-events.md. A signal is a lead for a reader, not a finding of intent.

Prices. Stated prices are read with driftgrade's price patterns, and a stated amount can be an example rather than a charge. On 16 servers a tool's single stated price became a different single price. Prices fell on 10 of them, rose on 3, and moved both ways on 3. One server carried 285 of the 414 before-and-after pairs (68.8%), so these are reported per server. 4 servers at least tripled a price. The most servers with such a readable change within four days was 15, from 2026-09-24. Counted more loosely, 55 servers had a change in the set of amounts an existing tool stated within four days from 2026-09-24; that count includes an amount added beside others or dropped from a list, so it is not a count of price rises. An earlier draft's claim of 23 servers in four days was withdrawn because it does not reproduce under the strict definition, a readable change together with a tripling.

Coverage. Of 19,764 hosted state files, 13,255 (67.1%) were read and current, and 369 (1.9%) had a catalogue but a failed latest attempt, 151 of them with HTTP 429. 6,140 (31.1%) had never been read: 4,165 (21.1%) answered HTTP 401, 402 or 403, which usually asks for a login or payment, though a 403 can be a block rather than a login; 586 (3.0%) answered 404, 773 (3.9%) failed at the network, 262 (1.3%) returned a 5xx, 248 (1.3%) another 4xx, and 98 (0.5%) failed the protocol.

Lookup. The log's lookup file held 230,763 distinct tool definitions, of which 217,985 had been seen on exactly one server. Definitions are compared byte for byte after canonicalisation (RFC 8785), so two that differ in nothing but a number, such as a ticking counter, count as distinct.

Earlier studies. The churn study's figures (docs/CHURN.md) reproduced from its committed data: across 109 third-party servers and 522 upgrades, a pin stopped on 235 (45%) and a description was reworded in 150 (29%). The scan study (docs/SCAN.md) read 240,464 tools from 13,170 servers on 2026-09-23; it is quoted as dated and was not rerun. Its finding was that the rules matched nothing a person reading the hits judged to be a poisoning payload. Its caveats stand: pattern rules and one reader, the semantic tier not run, and five of sixteen reworded attacks caught.

Limitations

  • What a server says, not what it does. A tool can keep an identical description while its code changes, and the log would not record that as a change (docs/TRANSPARENCY.md, "What it cannot do").
  • One reader, a few times a day. A version a server shows for a few hours between two readings can be missed (docs/TRANSPARENCY.md, "What it cannot do").
  • Signed-in clients see more. The log reads what an anonymous client is shown (same section).
  • A server that recognises the reader and shows it the same poisoned tool it shows a target is not caught by the log, though that tool is then on the public record (same section).
  • One publisher. The log is published by one party, not by several independent logs that watch each other (same section).
  • The seal bounds dates from above. It says nothing about whether the data is accurate, or that the log left nothing out (docs/TRANSPARENCY.md, "What it does not prove").
  • Records before 27 September 2026 are not sealed, and cannot be (docs/TRANSPARENCY.md, "From which date").
  • A baseline, not a trend. The observed window ran from 2026-09-23 to 2026-09-29.
  • Coverage. 6,140 of 19,764 hosted servers (31.1%) had never been read, with the reasons above.

Reproducing

git clone https://github.com/rufat325/heldfast-feed feed
git -C feed -c core.autocrlf=false archive 08a6e310855f6facb9d3bf6e4203db59733bd1ca \
  | tar -x -C feed-snapshot          # a case-sensitive directory on Windows or macOS
git clone https://github.com/rufat325/heldfast && cd heldfast
git checkout b9f7ec8c10367e74abb5044ac02aee86405236a0
python research/report/numbers.py --feed ../feed-snapshot --out research/report/numbers.json
python research/report/numbers.py --feed ../feed-snapshot --verify
python research/report/check_numbers.py docs/FIRST-WEEK.md

Checking the checkpoint itself is described in docs/TRANSPARENCY.md, "How to verify one". Some paths in the feed differ in nothing but case, so on Windows or macOS extract into a case-sensitive directory.

References

Appendix: claims register

Generated by research/report/numbers.py from numbers.json. Do not edit by hand.

C01

figure value denominator definition
npm watched 9712 npm packages on the watchlist.
hosted watched 19746 Hosted endpoints on the watchlist.
total watched 29458 Every entry on the watchlist.
hosted watched pct 67.0 entries on the watchlist (29458) Hosted endpoints' share of the watchlist. As a percentage, one decimal.
watchlist synced "2026-09-28" The date the watchlist was last rebuilt from the registry.

C02

figure value denominator definition
hosted state files 19764 Hosted servers the feed holds a state file for, including ones since dropped from the watchlist.
read and current 13255 Hosted state files with a catalogue and no failed attempt since.
read and current pct 67.1 hosted state files (19764) Read and current, of all hosted state files. As a percentage, one decimal.
catalogue but latest failed 369 Hosted state files with a catalogue whose latest reading attempt failed.
catalogue but latest failed pct 1.9 hosted state files (19764) The same, of all hosted state files. As a percentage, one decimal.
catalogue but latest failed by reason {"HTTP 429": 151, "HTTP 400": 35, "URLError": 30, "HTTP 404": 28, "TimeoutError": 28, "HTTP 401": 21, "HTTP 530": 15, "HTTP 503": 15, "HTTP 403": 9, "HTTP 402": 8, "HTTP 502": 8, "no tools/list answer": 7, "HTTP 410": 5, "HTTP 500": 4, "HTTP 405": 2, "HTTP 522": 2, "HTTP 409": 1} Those, by the recorded reason (attempted.why).
never read 6140 Hosted state files with no catalogue: never read.
never read pct 31.1 hosted state files (19764) Never read, of all hosted state files. As a percentage, one decimal.
never read by reason {"HTTP 401": 4057, "URLError": 757, "HTTP 404": 586, "HTTP 400": 110, "HTTP 503": 84, "no tools/list answer": 72, "HTTP 402": 69, "HTTP 502": 57, "HTTP 405": 56, "HTTP 500": 47, "HTTP 530": 45, "HTTP 403": 39, "HTTP 429": 39, "HTTP 410": 36, "initialize failed": 26, "HTTP 522": 17, "TimeoutError": 16, "HTTP 307": 8, "HTTP 521": 4, "HTTP 421": 4, "HTTP 526": 3, "HTTP 451": 2, "HTTP 525": 2, "HTTP 501": 2, "HTTP 422": 1, "HTTP 523": 1} Never read, by the recorded reason (attempted.why).
never read: login or payment 4165 Never read, reason grouped as login or payment.
never read: login or payment pct 21.1 hosted state files (19764) Never read for login or payment, of all hosted state files. As a percentage, one decimal.
never read: not found 586 Never read, reason grouped as not found.
never read: not found pct 3.0 hosted state files (19764) Never read for not found, of all hosted state files. As a percentage, one decimal.
never read: other 4xx 248 Never read, reason grouped as other 4xx.
never read: other 4xx pct 1.3 hosted state files (19764) Never read for other 4xx, of all hosted state files. As a percentage, one decimal.
never read: 5xx 262 Never read, reason grouped as 5xx.
never read: 5xx pct 1.3 hosted state files (19764) Never read for 5xx, of all hosted state files. As a percentage, one decimal.
never read: network error 773 Never read, reason grouped as network error.
never read: network error pct 3.9 hosted state files (19764) Never read for network error, of all hosted state files. As a percentage, one decimal.
never read: protocol 98 Never read, reason grouped as protocol.
never read: protocol pct 0.5 hosted state files (19764) Never read for protocol, of all hosted state files. As a percentage, one decimal.
never read: other 8 Never read, reason grouped as other.
never read: other pct 0.0 hosted state files (19764) Never read for other, of all hosted state files. As a percentage, one decimal.

C03

figure value denominator definition
events 13021 Every change event in the snapshot's events/*.jsonl.
hosted events 12642 Change events of hosted servers.
npm events 379 Change events of npm servers.
npm events seeded 246 npm events carried in from the churn study (seeded), not observed by the feed.
npm events live 133 npm events the feed observed.
live events 12775 Every event the feed observed (not seeded).
live window start "2026-09-23T04:23:42+00:00" The earliest observed_at among observed events.
live window end "2026-09-29T07:38:43+00:00" The latest observed_at among observed events.
live events per UTC day {"2026-09-23": 1308, "2026-09-24": 6051, "2026-09-25": 266, "2026-09-26": 2056, "2026-09-27": 2227, "2026-09-28": 491, "2026-09-29": 376} Observed events by the UTC date of observed_at.
live hosted events per UTC day {"2026-09-23": 1267, "2026-09-24": 6028, "2026-09-25": 257, "2026-09-26": 2041, "2026-09-27": 2214, "2026-09-28": 468, "2026-09-29": 367} Observed hosted events by the UTC date of observed_at.
live events per UTC day without top three {"2026-09-23": 182, "2026-09-24": 521, "2026-09-25": 262, "2026-09-26": 562, "2026-09-27": 599, "2026-09-28": 491, "2026-09-29": 370} The same, leaving out the three operators with the most hosted events.
busiest hosted day "2026-09-24" The UTC date with the most observed hosted events.
busiest hosted day events 6028 Observed hosted events on that date.

C04

figure value denominator definition
substantive 9722 Events of kind substantive (research/feed/operators.py: an event is substantive if a tool was added, removed or changed in substance).
numbers-only 3150 Events of kind numbers-only (research/feed/operators.py: an event is substantive if a tool was added, removed or changed in substance).
reorder-only 149 Events of kind reorder-only (research/feed/operators.py: an event is substantive if a tool was added, removed or changed in substance).
unclassified 0 Events of kind unclassified (research/feed/operators.py: an event is substantive if a tool was added, removed or changed in substance).
stats.json kinds {"numbers-only": 3150, "reorder-only": 149, "substantive": 9722} The kinds stats.json states.
recomputed kinds {"numbers-only": 3150, "reorder-only": 149, "substantive": 9722} The kinds recomputed from both catalogues of every event with research/feed/operators.py.
stats.json equals recomputed true Whether the two agree.

C05

figure value denominator definition
operators 1260 Distinct operators over every event: a hosted server's registrable domain, an npm package's name (approximate; see stats.json grouping).
top 1 operator "pipeworx.io" The operator with the 1. most hosted events.
top 1 events 7023 Hosted events of pipeworx.io.
top 1 events of hosted pct 55.6 hosted events (12642) pipeworx.io's share of hosted events. As a percentage, one decimal.
top 1 events of all pct 53.9 all events (13021) pipeworx.io's share of all events. As a percentage, one decimal.
top 1 kinds {"numbers-only": 2975, "substantive": 4048} pipeworx.io's events by kind.
top 1 events changing each field {"description": 7022, "inputSchema": 2527, "outputSchema": 1} pipeworx.io's events in which a changed tool's field moved, per field (one event counts once per field).
top 1 servers 1522 Distinct hosted servers of pipeworx.io with an event.
top 1 whole-catalogue events 0 pipeworx.io's events marked whole_catalogue (every tool changed at once).
top 1 whole-catalogue events pct 0.0 pipeworx.io's hosted events (7023) The same, of the operator's hosted events. As a percentage, one decimal.
top 2 operator "nolimit-observatory.workers.dev" The operator with the 2. most hosted events.
top 2 events 2341 Hosted events of nolimit-observatory.workers.dev.
top 2 events of hosted pct 18.5 hosted events (12642) nolimit-observatory.workers.dev's share of hosted events. As a percentage, one decimal.
top 2 events of all pct 18.0 all events (13021) nolimit-observatory.workers.dev's share of all events. As a percentage, one decimal.
top 2 kinds {"substantive": 2341} nolimit-observatory.workers.dev's events by kind.
top 2 events changing each field {"inputSchema": 2341} nolimit-observatory.workers.dev's events in which a changed tool's field moved, per field (one event counts once per field).
top 2 servers 2341 Distinct hosted servers of nolimit-observatory.workers.dev with an event.
top 2 whole-catalogue events 2341 nolimit-observatory.workers.dev's events marked whole_catalogue (every tool changed at once).
top 2 whole-catalogue events pct 100.0 nolimit-observatory.workers.dev's hosted events (2341) The same, of the operator's hosted events. As a percentage, one decimal.
top 3 operator "a2awire.com" The operator with the 3. most hosted events.
top 3 events 424 Hosted events of a2awire.com.
top 3 events of hosted pct 3.4 hosted events (12642) a2awire.com's share of hosted events. As a percentage, one decimal.
top 3 events of all pct 3.3 all events (13021) a2awire.com's share of all events. As a percentage, one decimal.
top 3 kinds {"substantive": 424} a2awire.com's events by kind.
top 3 events changing each field {"outputSchema": 424, "inputSchema": 106, "description": 47} a2awire.com's events in which a changed tool's field moved, per field (one event counts once per field).
top 3 servers 106 Distinct hosted servers of a2awire.com with an event.
top 3 whole-catalogue events 0 a2awire.com's events marked whole_catalogue (every tool changed at once).
top 3 whole-catalogue events pct 0.0 a2awire.com's hosted events (424) The same, of the operator's hosted events. As a percentage, one decimal.
top three events 9788 Hosted events of the top three operators together.
top three of hosted pct 77.4 hosted events (12642) The top three's share of hosted events. As a percentage, one decimal.
top three of all pct 75.2 all events (13021) The top three's share of all events. As a percentage, one decimal.
hosted servers changed 5257 Distinct hosted servers with at least one event.
median changes per changed hosted server 1 The median number of events per hosted server with at least one.
hosted servers with exactly one change 3107 Hosted servers with exactly one event.
changes per changed hosted server {"1": 3107, "2": 277, "3-5": 1789, "6-10": 61, "11-100": 23, "over 100": 0} Hosted servers with at least one event, by how many they had.
hosted events outside top three 2854 Hosted events of every other operator.
hosted servers outside top three 1288 Distinct hosted servers with an event, outside the top three operators.

C06

figure value denominator definition
review 66 Events the collector graded review: a change introduced a signal (heldfast.driftgrade).
quiet 12955 Events graded quiet.
review by introduced kind {"credential-path": 2, "critical-word": 2, "price": 58, "signal:concealment": 1, "signal:exfiltration": 2, "signal:mandated-side-effect": 2, "signal:override": 1} Review events per kind of signal introduced; an event with several kinds counts once under each.
price review events 58 Review events that introduced a price.
price review servers 41 Distinct servers with such an event.
first price review "2026-09-27T12:30:45+00:00" The earliest observed_at of a review event that introduced a price.
review without price 8 Review events that introduced no price; each is listed in research/report/review-events.md.

C07

figure value denominator definition
servers with a stated price changed 64 Servers with an observed event in which a stated price was changed (amounts found by heldfast.driftgrade's price patterns, per unit, in the description, title and schema text of a changed or added tool). A stated amount can be an example rather than the server's charge.
servers with a stated price added to a tool 45 Servers with an observed event in which a stated price was added to a tool (amounts found by heldfast.driftgrade's price patterns, per unit, in the description, title and schema text of a changed or added tool). A stated amount can be an example rather than the server's charge.
servers with a stated price removed from a tool 18 Servers with an observed event in which a stated price was removed from a tool (amounts found by heldfast.driftgrade's price patterns, per unit, in the description, title and schema text of a changed or added tool). A stated amount can be an example rather than the server's charge.
servers with a stated price on a new tool 82 Servers with an observed event in which a stated price was on a new tool (amounts found by heldfast.driftgrade's price patterns, per unit, in the description, title and schema text of a changed or added tool). A stated amount can be an example rather than the server's charge.
servers with a readable price change 16 Servers with at least one tool whose single stated price in a unit became a different single price in that unit.
readable price pairs 414 Such before-and-after pairs, per tool and event, over every server.
servers by direction {"both": 3, "down": 10, "up": 3} Servers with a readable change, by whether every pair rose, every pair fell, or both.
server with most pairs "remote/io.github.rccola990-cloud/x402-agent-store" The server carrying the most readable pairs.
pairs on that server 285 Readable pairs on that server.
pairs on that server pct 68.8 readable price pairs (414) That server's share of all readable pairs. As a percentage, one decimal.
most servers with a stated price changed in four days 55 The largest number of servers with a stated price changed (as above) within any 4 consecutive UTC days.
that four-day window starts "2026-09-24" The first day of that window.
most servers with a readable price change in four days 15 The same, counting only readable changes (one price became another).
that readable four-day window starts "2026-09-24" The first day of that window.
servers with a price at least tripled 4 Servers with a readable pair whose new price is three or more times the old.
claim of 23 servers in four days reproduced false Whether at least 23 servers had a readable price change within four days and at least one price at least tripled. Counting any change in stated amounts instead is the figure two above.

C08

figure value denominator definition
distinct definitions 230763 Entries in lookup/all.json: distinct tool definitions the log has recorded.
definitions on exactly one server 217985 Of those, the ones recorded on exactly one server.

C09

figure value denominator definition
checkpoints {"2026-09-27": [".json", ".json.ots"], "2026-09-28": [".json", ".json.ots", ".json.sigstore.json"]} Checkpoints in the snapshot tree, with the files each has (.json, .json.ots, .json.sigstore.json). The snapshot's own checkpoint is written after its commit, so it is not in its tree.

C10

figure value denominator definition
packages 109 Churn study, third-party servers: packages, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
release pairs 522 Churn study, third-party servers: release pairs, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
pin stops 235 Churn study, third-party servers: pin stops, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
pin stops pct 45 Churn study, third-party servers: pin stops pct, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
reword 150 Churn study, third-party servers: reword, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
reword pct 29 Churn study, third-party servers: reword pct, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
description edits 616 Churn study, third-party servers: description edits, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
introduced anything 25 Churn study, third-party servers: introduced anything, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
above 0.5 4 Churn study, third-party servers: above 0.5, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
above 0.5 releases 3 Churn study, third-party servers: above 0.5 releases, from analyse.py rerun (or suspicious.json, which needs results/ that are not committed to rerun).
matches docs/CHURN.md true Whether every figure above equals the one docs/CHURN.md states.

C11

figure value denominator definition
scan tools 240464 Tools the scan study read, quoted from docs/SCAN.md (data of 2026-09-23), not recomputed.
scan servers 13170 Servers it read, quoted.
scan hosted 12171 Of them hosted, quoted.
scan npm 999 Of them npm, quoted.
scan feed commit "87b7b05ace47" The heldfast-feed commit docs/SCAN.md names as its data.