Does anyone's tools/list actually paginate? A measurement.
Run: 2026-09-27 15:11 UTC · Sample: 40 servers · Servers that answered: 16 · Servers that paginated: 0
This exists because issue #177 was closed with an honest caveat: the fix walks a paginated tools/list to the end and refuses to present a truncated enumeration as complete, and that behaviour is executed by tests — but no test had ever observed a real third-party server paginate. The bounds in that fix (20 pages, 2,000 tools, dedupe by name, hard error on a repeated cursor) were therefore a stance about a protocol feature, not a survey of it. This page is the survey.
Method
tools/survey-mcp-tools-list-pagination.mjs reads https://registry.modelcontextprotocol.io/v0/servers (the official MCP registry), takes records that advertise a streamable-http remote, and sends each one:
- an anonymous
initializedeclaring protocol revision2025-06-18, - a
tools/listusing the returned session id when the server issued one, - a second
tools/listwith the returned cursor — only if the first response carried anextCursor.
No credentials are used or requested, nothing is written, and no server is called more than three times. Requests are sequential with a short delay.
The sample is not random. It is the first 40 streamable-http endpoints in the registry's default order at that timestamp, which is alphabetical by server identifier — so it over-represents names starting with a/ac/ad/ae/ag. Treat it as a snapshot, not as a population estimate.
Result
Of 40 attempted:
| Outcome | Count |
|---|---|
Answered initialize and returned a tool list | 16 |
Rejected anonymously (401) | 21 |
Rejected anonymously (403) | 1 |
400 on initialize | 1 |
502, no result | 1 |
The 16 that answered:
| Server | Endpoint | Tools on page 1 | Revision answered | Transport shape | nextCursor? |
|---|---|---|---|---|---|
ad.getle/leads | mcp.getle.ad/mcp | 35 | 2025-06-18 | JSON | no |
ai.afg/afg | afg.ai/mcp | 18 | 2025-06-18 | SSE | no |
ac.tandem/docs-mcp | tandem.ac/mcp | 13 | 2025-06-18 | JSON | no |
ai.agentberg/agentberg | agentberg.ai/mcp | 11 | 2025-06-18 | SSE | no |
agency.goji/goji | mcp.goji.agency/mcp | 9 | 2025-06-18 | SSE | no |
ai.adoraads/beauty | mcp.adoraads.ai/beauty | 6 | 2024-11-05 | JSON | no |
ai.agent-bev/bev-door | mcp.bev-buyer.ai/mcp | 6 | 2025-06-18 | JSON | no |
ag.hood/name-service | www.hood.ag/api/mcp | 5 | 2025-06-18 | JSON | no |
ai.advisorsai/service-navigator | advisorsai.ai/mcp | 5 | 2025-06-18 | JSON | no |
ai.afmr/discovery | afmr.ai/api/rpc | 4 | 2025-06-18 | JSON | no |
ai.afmr/discovery | afmr.ai/mcp | 4 | 2025-06-18 | JSON | no |
ad.inside/inside-ads | app.inside.ad/api/mcp | 3 | 2025-06-18 | JSON | no |
agency.ottobot/licensed-house-painters | house-painters-mcp.ottobot2025.workers.dev/mcp | 3 | 2025-06-18 | JSON | no |
agency.ottobot/contractor-license-changes | license-changes-mcp.ottobot2025.workers.dev/mcp | 2 | 2025-06-18 | JSON | no |
agency.ottobot/business-contact-finder | business-contact-finder-mcp.ottobot2025.workers.dev/mcp | 1 | 2025-06-18 | JSON | no |
ai.advisorsai/store-readiness | advisorsai.ai/store-readiness-mcp | 1 | 2025-06-18 | JSON | no |
Three things are worth more than the headline:
- 0 of 16 emitted
nextCursor. The largest list seen was 35 tools, returned in one page. So today the pagination path in our gateway is insurance against a server that has not been met yet — which is the honest way to describe a fix for an unobserved failure, and the way #177 should be read. - 1 of 16 answered with
2024-11-05after being asked with2025-06-18. A server in the wild does pick an older revision than the client declared. That is the concrete case for #178: when this was measured, our governed upstream client sent a revision and never read the answer, so a negotiation like this one was invisible to us and to its callers.masternow captures the answer and publishes it per upstream in theserversarray ofGET /mcp/tools- available from source and from the rolling:latest/:masterimage tags, which were built frommasteron 2026-09-27 - but not from the versioned0.8.0tag, which predates it - and whether a mismatching revision should be refused or merely reported is still being decided on the issue. - 22 of 40 advertised remote endpoints cannot be introspected anonymously (401/403). Any claim of the form "N% of MCP servers do X" that was measured this way is really a claim about the fraction that lets you look.
What this does not show
- It does not show that no MCP server paginates. 22 of the 40 refused the anonymous handshake (21x
401, 1x403) and 2 more failed beforetools/listwas reached, so 24 were never asked - and the sample is alphabetical, not random. - It does not show a size threshold. The spec gives no "paginate after N tools" rule, and the largest list observed here (35) is well under our 2,000-tool bound, so the bound is a guard against pathological answers, not a fitted parameter.
- It says nothing about stdio servers. This method can only reach advertised HTTP remotes; the stdio population is larger and is not addressed here.
- It is one timestamp. Servers change; a re-run is a new measurement, not a regression test.
Reproduce
node tools/survey-mcp-tools-list-pagination.mjs 40The script prints a JSON tally plus one row per server. It makes outbound requests to third-party services that are not ours; run it knowingly, and keep the timestamp with any number you quote from it.
Related measurement
The same 40 endpoints were asked a second question an hour later: will they agree to a protocol revision that does not exist? Two of the eighteen that produced a JSON-RPC answer echoed 9999-99-99 with HTTP 200, and one server answered 2024-11-05 to both probes. See mcp-protocol-revision-tolerance.md.