The State of the MCP Ecosystem 2026

The official MCP registry holds 18,843 unique servers across 60,374 published version rows, from 11,783 distinct namespaces. The earliest record is dated 8 September 2025; the most recent in this snapshot is 28 July 2026 at 10:29 UTC.

18,642 servers are active and 201 are marked deprecated, a deprecation rate of 1.1%. Deprecation is a publisher action, so this measures declared withdrawal rather than abandonment.

Every server carries a description, because the field is mandatory. The schema caps it at 100 characters: the longest description observed is exactly 100 characters and the median is 87.

Counting each server once in the month of its first published version, new servers rose from 456 in September 2025 to 4,571 in the first 28 days of July 2026. That is a 10.0× increase, and an average of 163 new servers per day in July.

Counting each server once in the month its current version was published produces a different and steeper curve, 352 to 6,133. This second curve measures recent publishing activity, not arrivals, because any server that shipped an update is re-dated into the month of that update. Only the first curve measures growth. Both are published in the dataset so the distinction can be checked.

July 2026 is incomplete. The snapshot was taken at 10:29 UTC on 28 July, so the final day is partial. Within the month, weekly arrivals ran 1,119, 1,274, 1,265 and 913. July should not be compared against complete months without that adjustment.

Across the whole registry, 8,359 servers are remote-only, 9,182 are local-only, and just 982 (5.2%) offer both. The ecosystem is two populations that barely overlap, and no monthly cohort has ever exceeded 6.1% dual-mode.

That split has now tipped. Local-only servers dominated every cohort from November 2025 to May 2026, peaking at 71.5% of February 2026 arrivals. In June 2026 remote-only overtook local-only for the first time, 53.1% against 40.2%, widening to 54.3% against 39.7% in July.

The practical consequence: tooling that assumes MCP servers are packages a user installs is now addressing the shrinking half of new supply. A client that cannot speak to a hosted endpoint cannot reach the majority of what shipped this quarter.inference

7,522 of 9,341 remote-capable servers (80.5%) declare no authentication header in their registry metadata. Measured per endpoint rather than per server, it is 7,843 of 9,744, also 80.5%. Of those that do declare one, 1,331 servers use an Authorization header and 275 use an API-key header.

This is not a finding that 80% of MCP servers are unauthenticated. It cannot be. A server that implements OAuth 2.0 with protected-resource metadata discovery, per RFC 9728, correctly declares no header: the client discovers the authorisation requirement at connection time. Those servers are fully authenticated and are counted here under none-declared. The same is true of servers using mutual TLS or network-level controls.

What the finding does establish is narrower and, we would argue, more useful: for four out of five remote MCP servers, the registry entry gives a client no way to determine the authentication requirement before opening a connection. Determining whether an endpoint needs credentials, and which, requires contacting it. That is a property of the registry schema, not a defect in any individual server.inference

A counterweight, and one we expected to come out worse: all 9,744 remote endpoints use HTTPS. Not one plaintext endpoint, not one loopback address, not one bare IP address appears in the registry. Transport security in declared MCP endpoints is, at this snapshot, universal.

The one hygiene exception is small: 53 servers publish endpoints on ephemeral development-tunnel domains, addresses that are not designed to survive as durable service URLs. Against 9,341 remote servers this is a rounding error, but it is a real category of stale entry.

Across current versions, 9,013 endpoints declare streamable HTTP against 731 declaring SSE. By server: 8,635 streamable-HTTP-only, 440 SSE-only, 266 supporting both.

The aggregate understates how fast this happened. Measured as the share of each cohort’s remote-capable arrivals that declare SSE, SSE held a 10.4% to 19.8% band in every month through May 2026, then fell to 2.2% in June and 2.5% in July. The transition was not gradual; it occurred within a single month.

One outlier is worth flagging rather than smoothing away. December 2025 reads 52.0%, but that cohort contained only 75 remote arrivals, so a single publisher moves the figure by tens of percentage points. It should not be read as a trend, and the June collapse holds with or without it.

Among the 10,164 package-based servers, npm accounts for 6,736 package declarations and PyPI for 2,927, a ratio of 2.3 to 1. The MCP bundle format (mcpb) records 684, OCI images 648, NuGet 88 and Cargo 11.

The bundle format is the one to watch. mcpb declarations among new arrivals held between 6 and 41 per month through May 2026, then rose to 103 in June and 94 in July, a 2.5× step against the preceding three-month level while npm grew 16% over the same span.

11,366 servers (60.3%) have published exactly one version and never shipped a second. 7,477 have published two or more.

The obvious objection is that the registry grew tenfold in ten months, so most servers are simply too new to have needed an update. The cohort data does not support that. The September 2025 cohort, ten months old at snapshot, sits at 59.6% single-version. Every cohort from September 2025 to July 2026 falls between 48.0% and 64.1%. Restricted to cohorts at least three months old, the rate is 57.9%. Single-version publication is a stable structural property of this registry, not a function of age.

A fairer counterweight is update recency, which is healthier than the single-version figure suggests: 13,191 servers (70.0%) had their current version published within the last 90 days, and only 1,212 (6.4%) have gone more than 180 days without a new version.

At the other extreme, publishing is dominated by automation. The 50 most frequently published servers account for 12.5% of all 60,374 version rows. Twenty-five servers have shipped 100 or more versions, and one has shipped 1,154. The version feed reflects continuous-integration pipelines more than it reflects human release decisions.

The largest namespace, io.github.pipeworx-io, holds 1,279 servers, 6.8% of the registry. The five largest hold 10.9% and the twenty-five largest hold 18.1%. Eight namespaces hold 100 or more servers each.

Reported alone, that overstates the case. 10,514 of 11,783 namespaces (89.2%) publish exactly one server, and the hundred largest namespaces together account for only 25.9% of the registry. Bulk publishing is visible at the head and almost absent across the body.

July’s growth is broad rather than bulk-driven. The month’s 4,571 new servers came from 3,415 distinct namespaces, with the ten largest contributing only 11.5%. Recent growth is not an artefact of a small number of automated uploads.

Of 18,843 servers: 15,345 (81.4%) link a source repository, 9,642 (51.2%) set a title, 8,122 (43.1%) give a website, and 1,013 (5.4%) provide icons. Only 564 servers (3.0%) supply all four.

Completeness is not trending upward. Icon adoption has moved within a 2.0% to 7.1% band across every monthly cohort since September 2025, with no directional improvement despite a tenfold increase in publishing volume. Discovery surfaces that depend on rich metadata will be working with a 3% to 5% subset for the foreseeable future.

Descriptions are near-universally unique, at 18,508 distinct strings across 18,843 servers, but the shared ones are revealing. 21 servers ship the literal scaffolding placeholder “Description of my MCP server”, 8 ship “An MCP server that provides [describe what your server does]”, and 578 servers have descriptions shorter than 40 characters.

Source hosting is effectively monocultural: 15,322 servers link to GitHub against 23 linking to GitLab, a ratio of 666 to 1. Namespace top-level domains follow the same shape, with io at 14,035, com at 2,188, ai at 863 and eu at 125.