State of the MCP Ecosystem, July 2026 · MCP Queen

State of the MCP Ecosystem — July 2026

MCP Queen probes every remote server in the official Model Context Protocol registry on a continuous cycle. This report is generated from 43,320 live probes of 9,326 remote servers , plus registry metadata for the full corpus. Every number is reproducible from the public API and the published dataset (CSV, 9,326 rows). Nothing here is self-reported.

The headline numbers

18,849 servers in the official MCP registry (18,650 marked active).

9,312 (49.9%) advertise a remote endpoint. The other 9,338 are local-install only (npx/uvx/docker): half the registry cannot be used without pulling code onto your machine.

Of the 9,326 remotes we grade , 7,723 (82.8%) are reachable right now. 1,603 (17.2%) are dead : they advertise an endpoint that does not answer. Roughly one in six "live" MCP servers is not.

Grades

GradeServersShare of graded<br>A4,95653.1%<br>B1431.5%<br>C1,50116.1%<br>D9189.8%<br>F1,80819.4%

Grading is deterministic: reachability 25, protocol compliance 15, tooling quality 35, latency 10, provenance 15, and every point carries the verbatim observation that earned it. 29.2% of graded servers land at D or F , almost entirely on dead endpoints, broken initialize handshakes, or empty tool catalogs.

Authentication

5,201 servers (55.8%) are open: no auth required.

1,882 (20.2%) are auth-gated and well behaved: they reject unauthenticated calls with proper protocol semantics.

640 (6.9%) are auth-bare: they demand credentials but fail the rejection handshake itself, so a client cannot even discover how to authenticate correctly.

2,522 servers (27.0%) are graded provisional because auth gates their tooling: their advertised capabilities cannot be independently verified.

Tooling: the strongest part of the ecosystem

Across servers that answer tools/list, we have cataloged 102,013 tools from 5,241 servers (about 19 tools per server). Quality among servers that respond is high: 99.7% of cataloged tools ship a typed input schema and 99.5% carry a real description . The ecosystem's problem is not tool quality. It is that a large minority of servers cannot be reached or introspected at all.

Latency

Median round-trip for a reachable server: 233 ms . Open servers respond faster on average than auth-gated ones (336 ms vs roughly 480 ms mean).

What this means

Listing is not verification. Half the registry is local-only, a sixth of the remotes are dead, and a quarter of graded servers hide their capabilities behind auth. A registry entry tells you a server existed once; a probe tells you it works now.

The floor is low but the ceiling is high. The A-grade majority shows that a well-behaved remote MCP server is not hard to run. The 29% D/F tail is operational neglect, not technical difficulty.

Agents need freshness. Grades here churn daily as servers die and recover. Point-in-time directories go stale in days; continuous probing is the only honest signal.

Method, rubric, and per-server evidence: the evidence registry. MCP Queen is built by Health AI, the team behind the Constat and Clarity evidence servers.