MCP Registry Census — 2026-08-19

Slug: mcp-ecosystem-growth-100-million-installs-2026 Track: sandbox-poc Date: 2026-08-19 Evidence: sandbox-executed

The last official MCP server count is the "10,000 active servers" figure published in the December 2025 Agentic AI Foundation announcement. No newer count has been published. Because the official MCP Registry exposes a public, unauthenticated read API, the current number does not have to be estimated — it can be counted. This note records that count so the figures in the article can be reproduced or refuted.

What was measured

Every server in the official registry, restricted to the current version of each, with the registry's own active status used to exclude deleted and deprecated entries. For each active server the run records whether it declares a remote endpoint, an installable package, both, or neither, plus the transport and package-registry types it declares.

Command

python3 scripts/mcp-registry-census.py \
  --out data/lab-runs/mcp-ecosystem-growth-100-million-installs-2026.registry.json

The script pages GET https://registry.modelcontextprotocol.io/v0/servers at limit=100 with version=latest, sleeping 150ms between pages, and follows metadata.nextCursor until the cursor is exhausted. No authentication, no write operations, no rate-limit overrides. The complete field in the output is true only when paging ended because the cursor ran out rather than because the page cap was hit; a truncated run is therefore visible in the artifact rather than silently rounded off.

Result

{
  "endpoint": "https://registry.modelcontextprotocol.io/v0/servers",
  "query": "version=latest",
  "complete": true,
  "pages_fetched": 227,
  "rows_returned": 22658,
  "latest_version_entries": 22658,
  "active_latest_entries": 22408,
  "distribution": {
    "remote_only": 10678,
    "package_only": 10180,
    "remote_and_package": 1183,
    "neither": 367,
    "any_remote": 11861
  },
  "remote_transport_counts": {
    "streamable-http": 11262,
    "sse": 1034
  },
  "package_registry_counts": {
    "npm": 7427,
    "pypi": 3250,
    "mcpb": 839,
    "oci": 726,
    "nuget": 99,
    "cargo": 29
  }
}

Full artifact, including the namespace concentration counts: data/lab-runs/mcp-ecosystem-growth-100-million-installs-2026.registry.json

Derived figures used in the article

Figure	Derivation
22,408 active servers	`active_latest_entries`
52.9% expose a remote endpoint	11,861 / 22,408
47.7% remote-only	10,678 / 22,408
45.4% package-only	10,180 / 22,408
91.6% of remote transports are `streamable-http`	11,262 / 12,296 declarations
60.0% npm, 26.3% PyPI	7,427 and 3,250 of 12,370 package declarations
One publisher holds 5.9%	`io.github.pipeworx-io` at 1,312 / 22,408
Top 15 namespaces hold 14.9%	3,348 / 22,408
2.2x the December 2025 official count	22,408 / 10,000

Limits of this evidence

The registry counts what publishers have registered, not what anyone runs. It exposes no install or download counts, so nothing here supports a "most popular server" ranking, and the article does not make one.

Entry counts are also inflated by bulk publishing. A single namespace accounts for 1,312 entries and the top fifteen for 14.9% of the registry, so the server total measures publishing activity more than it measures distinct maintained integrations. The 2.2x growth against December 2025 is a comparison between our count and an official count whose methodology was never published; it may not be measuring the same thing.

The registry is a preview release and its maintainers state that data resets may occur before general availability. A re-run on a later date will not necessarily reconcile with these numbers.

Servers declaring neither a remote endpoint nor a package (367, 1.6%) are counted as active but are not usable as published; they are left in the active total rather than quietly dropped.