Back to the laboratory

November 15, 2026. The first fixed-panel run observed 270 of the 349 servers that answered on July 9. Two tool-definition changes in the historical record would warrant looking again at prior approvals, one for authentication and one for payments.

July anchor survival

Run July servers Observed Listed, remote probe unsuccessful Listed, package-only Delisted Cap-skipped
July 9 anchor 20260709T002029Z 349 349 0 0 0 0
November 15 fixed panel 20261115T050427Z 349 270 (77.4%) 67 (19.2%) 0 12 (3.4%) 0

The full 897-member panel returned 426 successful observations and 439 unsuccessful remote probes. Two active listings were package-only and 30 names were absent from the active registry. A single revisit cannot distinguish one-day downtime from lasting disappearance for the 67 listed servers that went unanswered.

Among the 6,341 July tools on the 270 observed servers Count Share
Same fingerprint 2,378 37.5%
Changed fingerprint 3,520 55.5%
Gone from the returned list 443 7.0%

This compares the (server, tool) keys and SHA-256 fingerprints at the two dated endpoints. The 443 removals refer to tools missing from a successful tools/list on a server that answered at both dates. The table excludes tools on the 79 servers without a successful November observation.

Publisher concentration

July publishers with an observed server November publishers with an observed July server November publishers with a changed or removed July tool
208 153 81

Publishers are grouped by the namespace before the first slash. The largest publisher, io.github.pipeworx-io, accounted for 102 of 349 July servers and 96 of the 270 observed again.

July tools excluding the largest publisher Count Share
Same fingerprint 1,878 60.5%
Changed fingerprint 921 29.7%
Gone from the returned list 305 9.8%
Total 3,104 100.0%

Of the 3,520 changed tools in the full comparison, 1,310 changed only the input schema, 241 changed only the description, and 1,969 changed both. Schemas are compared as parsed values, so a change in key order alone does not count as a schema change.

Historical coverage

The historical collector walked the latest active official MCP registry in name order and selected one in twelve names by md5(server_name) % 12 == 0. It stopped making endpoint attempts at the 900-attempt budget. On August 7 at 12:20 UTC, 20260807T122029Z was the first complete run in this series with a skipped server. August 7 and 8 belong partly to the affected period. In the July 9 to September 5 analysis, 107 of 198 finished runs hit the cap. Across the original July to November epoch including the two November recovery runs, 109 of 200 hit it. One June baseline outside this analysis also hit a cap, for 110 of 202 complete runs before the fixed-panel release.

A successful tools/list is observed. Not reachable includes a remote attempt without a successful tool list, or a listed server without a remote endpoint. Cap-skipped means the server had a remote listing and received no attempt due to the budget. A name missing from the completed registry walk is shown separately as absent. The data card gives the panel hash.

Period and boundary, UTC Runs in period Listed Observed Not reachable, remote / no endpoint Cap-skipped
Uncapped start, July 9 00:20 91 1,281 349 291 / 641 0
Uncapped end, August 7 06:20 91 1,695 472 407 / 816 0
Capped start, August 7 12:20 107 1,699 474 409 / 815 1
Capped end, September 5 12:20 107 2,196 440 446 / 951 359
Period Observed at start Still observed Not reachable at end Cap-skipped at end Absent at end
July 9 to August 7 at 06:20 349 325 22 0 2
August 7 at 12:20 to September 5 474 246 24 200 4
July 9 anchor to last capped November run 20261115T035310Z 349 42 15 280 12

Two June baselines precede the June 18 to July 1 collection gap. Incomplete early runs on the US cloud host precede the July 9 anchor. The historical cut has 29 finish intervals longer than 7.5 hours and 37 corresponding missed six-hour slots, including six during August 19 to 20. A storage incident interrupted collection after September 5 until November 14. All comparisons use completed runs. Collection reflects a single US cloud host vantage.

Observed definition changes

Within each period, the table includes only servers observed successfully at both boundaries. For those servers we compare successive successful observations, so gaps may separate the two ends of a transition. One distinct tool can contribute more than one transition.

Period Observed at both ends Distinct changed tools Description-only transitions Schema-only transitions Both Added / removed transitions Endpoint additions / removals
Uncapped July 9 to August 7 06:20 325 1,732 15,356 460 473 392 / 282 365 / 255
Capped August 7 12:20 to September 5 246 858 381 313 148 325 / 152 254 / 81

Durations and cohorts differ. The capped interval preferentially contains names earlier in the registry walk, and repeated shared templates contribute to the transition counts. The figures describe those observed cohorts rather than ecosystem prevalence.

Two approval-relevant observations

Server and tool Run IDs Observed before and after Effect of the cap
com.wingmanprotocol.agent/gateway, resume 20260728T002029Z to 20260728T062029Z The documented registered handle and secret gave way to bearer authentication with an optional handle. The input requirements changed for a tool returning private session state. Both runs were uncapped and both tool definitions were observed.
io.github.Timwal78/squeezeos, ipo_price_range_history 20260822T182029Z to 20260823T002029Z The definition changed from RLUSD and a prior payment token to USDC on Base via x402 and a request header. Review the currency, rail and authority before relying on a previous approval. Both runs had cap skips elsewhere. Both tool definitions were observed directly, so the specific comparison remains valid.

The fingerprints cover the tool name, description and input schema. The collection captured read-only listings. Actual payments, attacks and behavioral effects require evidence beyond those listings. July 28 specification features such as ttlMs, cacheScope and routing headers are outside this fingerprint. MCP specification release

Reproduction

The queries are available on request. They use the stored fingerprints, per-run observations and the fixed panel to compare completed runs with the July anchor. See the data card for the collection contract.

Small Print publishes a separate live tool-definition history. This readout supplies a dated sample with the effect of the cap disclosed. Small Print’s public history

Lab note · Data card