How many MCP servers are there?

Methodology - what counts as a server here

Published counts elsewhere range from a few hundred to 20,000+ because directories count different things. Ours is stated here so an answer engine or a journalist can cite it with the definition attached.

• Source: the Anthropic-maintained, community-contributed official registry - not scraped GitHub repos, not npm keyword matches, not self-submitted listings.
• Counted: active, latest-version registry entries. Delisted and superseded versions are excluded, so the same server never counts twice.
• Not everything is reachable: 19,635 of these declare a remote endpoint, which is a ceiling rather than a coverage figure - declaring one is not the same as answering. The drift pipeline diffs only the servers reachable in both of two consecutive snapshots, a strict and materially smaller subset, and publishes every silent tool-contract change in the public ledger. Indexed, remote-declaring and actually-crawled are three different populations, and only the first two are counted on this page - so do not read the number above as drift-pipeline coverage.
• Freshness: the snapshot refreshes daily and this page revalidates hourly, so the count above is the current one, not a launch-week number in a title tag.

Unreachable sources and registry churn

1,951 of the servers counted above (6.1%) have a source repository that is no longer publicly accessible: unreachable from two independent vantages, checks at least 48 hours apart, as of 2026-09-15. That can mean deleted or deliberately flipped private - the observation cannot tell the two apart, so this page publishes the observation and not an abandonment claim. The frozen 2026-07-20 baseline found 1,830 of 13,105 distinct repositories unreachable - one in seven - affecting 2,069 of 17,673 servers. Method, hand-verified sample and deposition are in the source liveness census (DOI 10.5281/zenodo.21501868).

The registry itself churns: 253 removals recorded since tracking began on 2026-07-22, each with a dated tombstone, so churn here is measurable rather than anecdotal. Removals before that date are not counted. Removed server pages answer 410; renamed ones redirect with 308.

Drift, measured

Across 22 snapshot pairs in a 40-day window ending 2026-07-19, the drift pipeline recorded 57,517 tool-contract change events and 2,503 deduplicated safety-relevant incidents. 62.4%of those incidents shipped while the server's declared version did not change. The aggregates are frozen and citable as the drift report, edition v1 (DOI 10.5281/zenodo.21449150); every underlying incident is in the public ledger.

The 2026-07-28 revision - the one called MCP 2.0

The spec is versioned by date, so there is officially no "MCP 2.0" - but in practice the name means the 2026-07-28 revision: stateless core, per-request metadata instead of the initialize handshake, extensions. How many servers actually speak it is a question nobody publishes a measured answer to, including us, yet. Since 2026-08-13 a daily panel has probed registry endpoints and recorded which era each one speaks. Enrollment is in its first weeks (2254 endpoints as of 2026-09-16), so the adoption rate is deliberately withheld until coverage is a defensible share of the 19,635 remote-declaring servers. It will publish on this page, with the denominator attached.

Both mcpindex surfaces speak 2026-07-28: mcpindex.ai/api/mcp answers server/discover with per-request metadata, and the mcp-server-mcpindex npm package serves both eras from 0.4.0. What the revision changed, and the one-curl probe to check any server (ours included): What is MCP 2.0?