MCP registry triples to 45,000 server names in two weeks, but a quarter of new endpoints are dead

The registry tripled in two weeks

The official MCP registry grew from 16,993 distinct server names on 3 October 2026 to 45,397 on 17 October, according to an independent census published on dev.to by Pennyforge, a one-person studio. Counted as name:version pairs, the registry expanded from 53,068 to 156,994 over the same 14 days.

Roughly 28,557 server names that had never existed before appeared, while only 256 pairs (0.48%) vanished. The author flags one caveat on the numbers: the registry is exposed through a preview API, and the 3 October walk died mid-list on an HTTP 500 before reaching the tail. That makes "+24.5% at least" the guaranteed floor and 2.96x the upper bound. Either way, with so few removals, the pattern is accretion rather than a data reset, which would show mass removal and re-addition.

Probing the newcomers

Registry metadata says nothing about which protocol a server actually speaks, so Pennyforge sampled 300 brand-new names — all absent from the 3 October snapshot — out of the 17,255 new registrations that expose a public streamable-HTTP endpoint. Each endpoint received real initialize handshakes for four dated protocol eras, two fictional eras (1999 and 2099) to catch servers that echo the version field without reading it, and the spec's own server/discover request. No tools were invoked, and the probe tool plus per-endpoint results are published alongside the post.

The results:

• 67 servers (22.3%) served the newest era, the July 2026 revision dated 2026-07-28
• 135 (45.0%) answered only 2025-era wire: 68 on 2025-11-25, 47 on 2025-03-26 and 20 on 2025-06-18
• 81 (27.0%) answered no handshake at all — cloud dev deployments, template URLs with literal placeholder braces, side projects abandoned mid-registration
• 12 (4.0%) blindly echoed every claimed era, including the fictional ones
• 5 (1.7%) failed with probe or transport errors

A migration that is not accelerating

Per the analysis, the July 2026 revision of the MCP spec is the first protocol change that can break clients: on the modern stateless HTTP path the initialize handshake is gone, and servers must implement server/discover. Pennyforge had measured new-era adoption at two earlier points, both before the influx: 9.7% of 186 endpoints on 1 October, and 24.8% of a 4,673-endpoint cohort on 8 October.

The new wave came in at 22.3% (95% confidence interval 17.6–27.0), statistically overlapping the earlier cohort. The flood of new registrations replicates the existing distribution rather than pushing the ecosystem toward the newer spec. Of the 67 modern servers, 37 negotiate the 2026-07-28 era through initialize (dual-era) and 30 are stateless-modern, answering discover directly.

Caveats on the measurement

The findings come from a single snapshot, with n=300 and a confidence interval of roughly ±4.7 percentage points. Only public streamable-HTTP endpoints were probed; stdio-only servers, which run inside the caller's own SDK, are excluded — so true ecosystem adoption is probably higher than this hosted-surface measurement suggests. An endpoint counted as dead may have worked at registration and lapsed days or weeks later. Endpoints also move in both directions: the author watched one server lose new-era support between two probes a week apart.

Why it matters

A client pinned to the 2026-07-28 era can use only about a quarter of new servers out of the box; three of four live new registrations still speak pre-2026 wire. Anyone auto-onboarding servers from the registry should budget 25–30% dead links among new entries, because registry status says nothing about liveness — and liveness is cheap to probe and expensive to discover in production. Finally, because 4% of endpoints echo back any version a client claims, the era a server advertises in its handshake is a claim, not evidence. The tripling is a genuine growth signal, but it measures quantity, not protocol maturity: new-spec adoption has held flat at roughly one server in five through the entire influx.