Skip to content

MCP Ecosystem Growth: 47,000 Servers in One Year

industry By AI Hub Team · updated

A year ago, Anthropic’s own marketing number for the Model Context Protocol was “10,000+ active public servers” (December 2025). As of October 21, 2026, the official MCP Registry lists 47,267 active servers — a ~4.7× increase in ten months, and a number that no longer needs the “promising new standard” qualifier.

MCP is now the default answer to “how does an AI application connect to the outside world,” and the growth curve says the ecosystem is in its adolescent-to-adult phase: governance has moved to the Linux Foundation, the SDKs crossed 500 million monthly downloads, and even the CVE process has caught up.

From one lab’s trick to a Linux Foundation standard

MCP was announced by Anthropic in November 2024 with a framing that stuck: “USB-C for AI applications” — one standard instead of a custom integration per data source, tool or workflow. It shipped open-source, which is the reason it spread past Claude at all.

The governance step happened on December 9, 2025, when the Linux Foundation launched the Agentic AI Foundation (AAIF) with three founding projects: MCP (Anthropic), goose (Block) and AGENTS.md (OpenAI). The foundation is co-founded by Anthropic, Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare and Bloomberg. Anthropic’s announcement framed the donation as the protocol outgrowing a single company’s roadmap — and the membership numbers back that up: 286 member organizations as of October 21, 2026, up from 49 at launch.

Microsoft is not a founding member, but it is a first-class participant: Microsoft and Anthropic ship an official C# SDK for MCP, which matters for the large enterprise .NET install base that the TypeScript and Python worlds do not reach.

The adoption numbers (as of October 2026)

MetricValueWhen
Active servers in the official registry47,267Oct 21, 2026
SDK downloads, TypeScript (npm)231.9M (Sep 2026) vs 31.3M (Sep 2025)7.4× YoY
SDK downloads, Python (PyPI)260.8M30 days to Oct 20, 2026
MCP project-wide SDK downloads~500M monthlyJul 2026
GitHub repos tagged mcp-server38,336Oct 2026
Stars on modelcontextprotocol/servers91,449Oct 2026
Claude marketplace connectors (MCP-based)995Oct 2026 (75+ in Dec 2025)

Two of the most-starred servers show where the demand actually is: Playwright MCP (38,472 stars) for browser automation and GitHub’s own MCP server (33,828 stars) for code hosting. If you are picking one server to start with, those two are the highest-signal options.

The registry mix also shifted direction. In July 2026, 31% of registered servers were remote-hosted; by October 21, 2026, that figure is 66%. MCP started as a local-process protocol and is becoming a remote API surface. The remaining 34% are local: 11,423 npm packages, 8,199 Docker images and 3,251 Python packages in the registry.

What changed technically in 2026

The official documentation now tracks a spec version 2026-07-28 (the previous dated specs: 2025-06-18, 2025-03-26, 2024-11-05), with a public registry at registry.modelcontextprotocol.io and a SEP (Specification Enhancement Proposal) process for extensions.

The SDK 2.0 line is what most builders actually felt:

  • TypeScript SDK 2.0.0 (April 14, 2026) implements protocol v2026-03-27 in full: structured tool output, streamable HTTP sessions, elicit sampling.
  • Python SDK 2.0.0 followed on April 8, 2026.

Client support is broad: the official docs name Claude, ChatGPT, Visual Studio Code, Cursor and MCPJam among MCP-capable clients — the “build once, integrate everywhere” claim is now testable in production, not a roadmap promise.

The caveat: 382 CVEs

Growth brought a security surface. The NVD now lists 382 CVEs that mention MCP as of October 21, 2026. Most are in third-party servers, not the protocol itself — but the pattern is the same as every plugin ecosystem before it: the protocol is safe, the long tail is not. Practical rule: prefer official and well-maintained servers, treat every remote server as untrusted code with credentials, and audit before you grant a server access to anything sensitive.

What we recommend

  • Start with one or two official servers — Playwright or GitHub — and wire them into a real workflow before adding more. A dozen untested community servers are a security review, not a stack.
  • If you run agents on a machine you control, pair MCP with a local runner; the LM Studio vs Ollama comparison covers which runner’s MCP support fits your setup.
  • If you are building an agent product, MCP is now the integration layer you build against — the registry and AAIF governance mean it will outlive any single vendor’s roadmap.
  • Learn the protocol hands-on in our What is MCP: A Practical Guide, and browse the server directory in the Agents & MCP pillar.

Sources