Prolepsia forecastYes / no
Will the official MCP Registry list more than 10,000 servers on 31 December 2027?
As of 17 November 2026, Prolepsia puts the chance of yes at 99%.
99%chance of yes
How Prolepsia reasons
The report Prolepsia wrote with this forecast.
How forecasts are made · Track record
TL;DR
The forecast is 98.9% Yes and 1.1% No for more than 10,000 distinct publicly listed servers on December 31, 2027, at 23:59 UTC. The registry already stands far above the threshold; the question is whether that catalog survives, not whether growth continues.
Key evidence
- A large existing buffer. MCP Registry Observatory reported 66,040 listings on November 17, 2026, measured at 15:49 UTC and synchronized through 13:07 UTC. Using that count as a proxy, reaching 10,000 requires an 84.86% net contraction without additions. Observatory’s repository deduplication and incremental deletion tracking mean its total is not an exact canonical-name census. (Observatory)
- Canonical-name counts support the same scale. Reported complete enumerations of the official listing endpoint returned 66,183, 66,184 and 66,188 distinct public names in nearby collection windows. These were moving observations, not atomic snapshots, but the differences are immaterial to the threshold. (Official listing API)
- The counting rule is broader than operational usefulness. Latest-version filtering avoids counting multiple versions separately. Deprecated entries can remain public, while deleted entries are excluded by default. A broken or inactive implementation does not automatically cease to count. (Official API documentation)
- Policy and maintenance favor continuity. Current moderation generally retains low-quality or overlapping implementations. October 23 maintainer comments distinguish revalidation from automatic takedowns, and the November 10 release addresses synchronization and production scaling. (Moderation policy; revalidation discussion; release)
Uncertainties
The strongest adverse evidence is the explicit warning that preview-stage data resets may occur. Spam and non-functioning entries can also be removed, so today’s catalog is not a permanent floor. (Official policy)
The decisive uncertainty is whether a reset, restrictive policy change, or shutdown removes most public listings before resolution. There is no well-matched historical failure-rate sample to quantify those risks. Temporary API downtime alone does not imply No: the resolution rules allow qualifying archived evidence and an official catalog-preserving successor.
Read the full report
TL;DR
The forecast is 98.9% Yes and 1.1% No for more than 10,000 distinct publicly listed MCP servers on December 31, 2027, at 23:59 UTC. The existing catalog gives the Yes outcome a large cushion, with November measurements around 66,000 listings. Treat this as a forecast of catalog preservation, not continued rapid growth.
Context
The resolution counts distinct public canonical server names in the official registry. Multiple versions of one name count once, while distinct names count separately even when they share an implementation or endpoint. The official API supports latest-version filtering and cursor pagination; deprecated entries can remain public, while deleted entries are excluded by default. These distinctions prevent version totals from being confused with the number of eligible servers. (Official API documentation)
The cutoff is December 31, 2027, at 23:59 UTC. Exactly 10,000 names means No; at least 10,001 means Yes. The question does not require listed implementations to work or remain maintained. It also permits a qualifying official successor and archived evidence during temporary API unavailability. Permanent shutdown without a catalog-preserving successor means No. Those rules make institutional continuity and listing policy more important than endpoint reliability.
Evidence
The historical backbone is the registry’s observed size, not an estimated industry-wide survival rate. There is no sufficiently comparable historical sample of official AI-protocol registries from which to estimate a reliable shutdown or reset frequency. The available catalog observations nevertheless establish a clear fact: this is not a registry approaching the threshold from below.
The dated monitoring histories are shown below. They are separate series with different methods, not a single harmonized census. The October 15 and October 18 Observatory values are implied by its reported changes between endpoints. (MCP Census; Observatory)
| Series | Observation date in 2026 | Reported or implied count |
|---|---|---|
| MCP Census | July 8 | 15,494 |
| Observatory | October 15 | 43,870 |
| MCP Census | October 17 | 45,286 |
| Observatory | October 18 | 45,635 |
| Observatory | November 14 | 63,992 |
| MCP Census | November 16 | 65,340 |
| Observatory | November 17, measured 03:49 UTC | 65,720 |
| Observatory | November 17, measured 15:49 UTC | 66,040 |
A separate research preprint reports an August snapshot with 21,643 servers and 72,606 version records. That distinction matters: version records are not distinct servers. The abstract does not give the exact August snapshot date, so the finding is a historical cross-check rather than a precisely dated point in the monitoring series. (Research preprint)
Another November 16 measurement reported 65,560 distinct names across 40,299 namespaces. It retained one current record per name. Its total differs from the other monitors, but it supports the same conclusion about scale. (Tanod)
Reported complete enumerations of the official public API provide more directly relevant count evidence. Two nearby November 17 collection windows returned 66,183 and 66,184 distinct public names. Another returned 66,188; a later enumeration that also inspected deleted records found 66,200 non-deleted names. These observations were collected over intervals, not at a single instant. They support the approximate catalog size without establishing an exact historical count at the forecast timestamp. (Official listing API)
The monitoring totals need qualification. Observatory collapses some entries associated with the same repository and tracks deletions incrementally. Repository deduplication can lower its count relative to canonical-name counting, while missed deletion updates can raise it. Its latest 66,040 figure was measured on November 17 at 15:49 UTC, with synchronization through 13:07 UTC. It is a useful size proxy, not an exact resolution-compatible census. (Observatory methodology and observations)
The arithmetic is decisive despite these differences. Starting from the Observatory proxy, reaching 10,000 requires an 84.86% net contraction without additions. An 80% reduction still leaves approximately 13,208 listings. These are stress tests, not guaranteed floors. They show why flat growth, slower adoption, or a moderate cleanup does not by itself produce No. (Observatory baseline)
Concentration is the next question. A catalog dominated by a handful of publishers would be more exposed to targeted removals. Reported namespace analysis found 37,058 namespaces containing exactly one name. The five largest namespaces together held 8,473 names; removing all five from the 66,183-name observation leaves 57,710. Namespace diversity does not establish independent ownership, but it weakens the idea that removing a few conspicuous bulk publishers would erase the cushion. (Official listing data)
Cleanup pressure is real. An October 11 moderation issue reported approximately 2,365 near-identical entries under one namespace. That is evidence of bulk publishing and a reason to expect scrutiny. It is not evidence that those entries were removed, or that the rest of the catalog has the same composition. (Moderation report)
Current policy is permissive. It generally retains low-quality, buggy, vulnerable, or functionally similar servers, while allowing removal of spam and non-functioning entries. The same documentation explicitly warns that preview-stage data resets may occur. The first point supports persistence under routine moderation; the second creates a distinct risk that ordinary deletion statistics cannot capture. (Official moderation policy)
Recent decisions reinforce the distinction between quality problems and delisting. The November 6 release explicitly permits different names to share a remote URL. October 23 maintainer comments distinguish routine revalidation from automatic takedowns. Under that approach, shared endpoints and ordinary implementation drift do not automatically remove public names. (November 6 release; revalidation discussion)
Maintenance also supports continuity. The November 10 release addresses synchronization performance and production deployment safeguards. That is evidence of ongoing investment in operating the service, not a promise to preserve every listing through the cutoff. The working-group charter assigns responsibility for operating the registry and expanding coverage, while the preview documentation provides no uptime or data-durability guarantee. (November 10 release; working-group charter; preview documentation)
What's non-obvious
This is a preservation question, not a growth question. Rapid recent expansion is supportive, but the forecast does not require it to continue. The catalog can stagnate or shrink substantially and still clear the threshold. Extrapolating the recent growth rate would add apparent precision without addressing the main failure modes.
The second distinction is between ecosystem health and the resolving count. A server can stop working yet remain publicly listed. Conversely, a reset can remove listings even if the underlying implementations survive elsewhere. Latest-version publication dates also do not establish first-registration dates: an old name can publish a new version. The useful evidence is the stock of distinct public names and the rules governing their removal, not raw publication activity or the number of working endpoints. (API documentation)
Uncertainties
The largest uncertainty is a discontinuous policy or institutional event: a destructive reset, mandatory republishing that loses most entries, sweeping eligibility restrictions, or permanent shutdown without a qualifying successor. The preview warning makes reset risk concrete, but it does not establish that a reset is scheduled. Active maintenance supports continuity without guaranteeing it. (Moderation policy; releases)
Measurement uncertainty remains smaller but real. The full enumerations were not atomic snapshots, monitoring methods differ, and namespace diversity is not verified ownership diversity. A timestamped canonical-name export with deletion status, followed by comparable snapshots, would improve the count evidence. A clear policy on preview resets and catalog preservation would address the more consequential uncertainty.
The saved forecast remains 98.9% Yes and 1.1% No. The large current buffer explains the strong Yes lean; reset, policy, and shutdown risks explain why it is not certainty.
Forecast history
| Date | Forecast |
|---|---|
| 99% chance of yes |
Sources Prolepsia read
- Official API documentationraw.githubusercontent.com
- MCP server statistics (November 2026): how many MCP servers exist — The MCP Censusmcpcensus.com
- MCP Registry Observatory — health of the MCP registriesmcptoplist.com
- [2609.14119] Same Name, Different Server: A Security Census of Silent Drift in the Model Context Protocol Ecosystemarxiv.org
- State of the MCP Registry: 65,560 servers, remotes, packages and growth | Tanodtanod.dev
- Official listing APIregistry.modelcontextprotocol.io
- Moderation: ~2,365 near-identical servers registered under `io.github.sadri-dridi/*` · Issue #1701 · modelcontextprotocol/registry · GitHubgithub.com
- MCP registry changes on 2026-11-16 · ProtocolProbeprotocolprobe.com
- Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.mdwww.linuxfoundation.org
The question
This question asks whether the official Model Context Protocol (MCP) Registry will contain more than 10,000 distinct, publicly listed MCP servers on December 31, 2027. The official MCP Registry, hosted at https://registry.modelcontextprotocol.io, is a community-driven directory for discovering and publishing MCP servers. Its public API provides paginated access to registered servers and supports retrieving the latest published version of each server. As of November 17, 2026, the registry has already substantially exceeded the proposed threshold. Independent monitoring by MCP Registry Observatory reports approximately 65,720 server records in the official registry, while other registry-monitoring services report somewhat different totals depending on their counting methodologies. The threshold of 10,000 distinct servers has therefore already been surpassed, although the registry's size could change before the resolution date. The forecast concerns whether the registry still lists more than 10,000 distinct servers on December 31, 2027, rather than whether it reaches that milestone at any point before then. The count will include all publicly listed server entries regardless of whether their implementations are operational, actively maintained, or commercially supported. Sources: Official MCP Registry API documentation (https://registry.modelcontextprotocol.io/docs), MCP Registry Observatory (https://mcptoplist.com/observatory), and official registry API specification (https://github.com/modelcontextprotocol/registry/blob/main/docs/reference/api/generic-registry-api.md).
How it resolves
The question resolves YES if the official MCP Registry contains at least 10,001 distinct publicly listed MCP servers on December 31, 2027, at 23:59 UTC. It resolves NO if the registry contains 10,000 or fewer distinct publicly listed MCP servers at that time. The primary resolution source will be the official MCP Registry's public API at https://registry.modelcontextprotocol.io. The count will be determined by querying the server-listing endpoint (currently GET /v0.1/servers), retrieving all available pages, and counting distinct server identifiers using the registry's canonical server-name field. Where supported, the API's version=latest filter will be used to retrieve a single entry per server. Otherwise, entries will be deduplicated by canonical server name so that multiple published versions of the same server count as one server. Both remotely hosted and locally installed MCP server implementations will count if they have publicly visible registry entries. Entries removed, unpublished, or otherwise excluded from the registry's public listings by the resolution time will not count. The registry's own public records will take precedence over third-party estimates. If the API is temporarily unavailable at the specified time, an independently archived snapshot demonstrably representing the registry's state on December 31, 2027, will be used where available. Eligible evidence includes timestamped registry exports, archived API responses, or independently maintained complete snapshots of the official registry. If an exact snapshot is unavailable, the closest verifiable snapshot from December 31, 2027, may be used, provided there is no credible evidence that the count crossed the 10,000-server threshold between the snapshot and 23:59 UTC. If the registry has permanently shut down before the resolution date without an official successor preserving its listings, the question resolves NO because the registry no longer publicly lists more than 10,000 servers. If the official registry has migrated to a successor service, that service will be used if it clearly preserves the official registry's function and server catalog. If the registry remains operational but no sufficiently reliable evidence of its server count can be obtained, and the available evidence cannot establish which side of the threshold the count falls on, the question will be annulled rather than assigned an unsupported outcome. Sources: Official MCP Registry (https://registry.modelcontextprotocol.io) and official registry API documentation (https://github.com/modelcontextprotocol/registry/blob/main/docs/reference/api/official-registry-api.md).
Fine print
The threshold is strictly greater than 10,000: exactly 10,000 distinct servers produces a NO resolution, while 10,001 produces YES. A server is counted once per unique canonical registry name, regardless of how many versions, packages, transport configurations, deployment endpoints, or tools it provides. Distinct registry names count separately even if their underlying implementations are similar or identical. Published registry entries are counted without independently checking whether their associated repositories, packages, or network endpoints remain accessible. The count excludes servers listed exclusively in third-party MCP directories, GitHub repositories, package registries, or other catalogs unless they also appear in the official MCP Registry. Entries that are pending publication, rejected, or not publicly discoverable are excluded. If the registry changes its API structure or identifier format, the equivalent public listing and unique server identity under the replacement system will be used. The count is a snapshot of the registry on the specified date, not a cumulative count of every server ever registered. All dates and the resolution cutoff use UTC. Because the threshold had already been exceeded by November 2026, this forecast primarily concerns whether the registry maintains more than 10,000 distinct listings through the end of 2027, including the possibility of substantial removals, policy changes, or registry discontinuation.
More public forecasts
Will China's official defence budget for 2027 grow by more than 7%?
57%chance of yes
Prolepsia forecast
Will China raise its official budget deficit target above 4% of GDP at the March 2027 National People's Congress?
36%chance of yes
Prolepsia forecast
Will NOAA's Climate Prediction Center report a three-month Niño 3.4 index value (ONI or relative ONI, whichever CPC uses as its official index) for any season from SON 2026 to MAM 2027 higher than the peak of the 2015-16 El Niño on the same index?
84%chance of yes
Prolepsia forecast
Will China's official real GDP growth for 2027 be 4.5% or higher?
53%chance of yes
Prolepsia forecast
Will the US Commerce Department add DeepSeek to its Entity List before 1 January 2028?
55%chance of yes
Prolepsia forecast
Will Microsoft raise the list price of Microsoft 365 Copilot for business users before 31 December 2027?
17%chance of yes
Prolepsia forecast