When Anthropic first announced plans for an official MCP registry in early 2025, developers expected something like npm or PyPI for agent tooling: a curated, searchable index of servers that met a quality bar. What shipped was different. The official MCP registry is live, growing, and now holds close to 10,000 server records. But “official” and “curated” are not the same thing.

What the Official Registry Actually Is

The registry launched as a community-driven project under the Agentic AI Foundation, a directed fund within the Linux Foundation. Anthropic donated MCP to the foundation in late 2025, and the registry came with it. The code is open source on GitHub, and anyone can submit a server.

That last part is key. The registry is not an approval process. It is a feed. Servers are listed if they conform to the MCP specification and include basic metadata: name, description, transport type, and a valid package reference. There is no manual review for code quality, no verification that the server works reliably, and no scoring of documentation or install experience.

As of late May 2026, the registry API returned 9,652 server records. That number has likely crossed 10,000 by now. For comparison, the broader MCP universe (scraped from GitHub, npm, and third-party indexes) exceeds 48,000 packages that mention MCP. The registry captures roughly 20% of the total surface area.

What Made the Cut

Getting into the registry requires meeting the MCP specification. In practice, that means:

  • Valid transport declaration. The server must declare stdio, SSE, or HTTP transport. Servers without a transport type are rejected.
  • Package metadata. A name, version, and description. The registry pulls this from the package manifest.
  • Spec compliance. The server must implement the MCP handshake and expose at least one tool or resource.

That is it. If your server starts, responds to an initialize call, and has a valid manifest, it gets listed. The barrier is low by design. The foundation’s position is that discoverability should be open, and quality filtering should happen at the client or directory level.

What Didn’t Make the Cut

The registry rejects servers that fail spec compliance checks. Servers that crash on initialization, that return malformed tool schemas, or that do not implement the required protocol handshake are excluded. But the rejection criteria stop there.

Servers the registry does not filter out:

  • Abandoned projects. A server published once and never updated is listed indefinitely. There is no staleness check.
  • Duplicate functionality. Dozens of “weather API” or “hello world” servers sit alongside production-grade integrations. The registry does not deduplicate.
  • Missing or broken auth. Servers that require API keys but do not document how to configure them are listed like any other.
  • Minimal documentation. A one-line description is enough. Full guides, usage examples, and troubleshooting are not required.

The result is a registry that functions more like a package index than a curated directory. Useful for machines and automated discovery. Less useful for a developer trying to find the right tool for a job.

Where Curation Happens Instead

The open design of the official registry created a natural gap: builders need quality signals that the registry does not provide. That gap is where curated directories come in.

AgentNDX maintains a hand-reviewed directory of MCP servers, each scored for documentation quality, install reliability, auth clarity, and real-world use cases. The directory currently tracks over 600 curated servers with verified install commands and category tags. The difference between 600 curated entries and 10,000 raw listings is not a rounding error. Those two numbers answer different questions: “does this server exist?” versus “should I use this server?”

Other projects fill similar roles. Smithery offers a searchable registry with community ratings. mcp.so aggregates servers with install instructions. Each applies a different quality lens to the same pool of servers.

The layered model is intentional. The official registry is the canonical feed. Curated directories are the filter. MCP clients like Claude Desktop and Cursor can pull from either layer depending on what their users need.

The Enterprise Auth Question

One gap the registry is starting to address: authorization. The MCP specification recently added an enterprise authorization layer, giving organizations a way to scope what agents can access and under what conditions. This matters because the registry lists servers that handle sensitive operations (database writes, cloud deployments, payment processing) alongside read-only informational tools.

Before the auth layer, every server managed its own auth model. Some used API keys. Some used OAuth. Some used nothing. The new layer does not mandate a specific auth method, but it gives clients a standard way to discover and enforce auth requirements. For the registry, this means server metadata can now include structured auth declarations that clients can inspect before connecting.

It is early. Most servers in the registry have not adopted the new auth metadata yet. But the direction is clear: auth is becoming a first-class dimension of server discovery, and the registry will get more machine-readable as a result.

What This Means for Builders

If you are building with MCP servers today, the practical advice has not changed much:

Do not browse the official registry like an app store. It is a feed, not a recommendation engine. Use it for automated discovery or when you already know the server name you want.

Use a curated directory for evaluation. When you need the best server for a specific use case, a quality-scored directory saves real time. Check whether the server has a verified install, documented auth, and recent updates before wiring it into a workflow.

Watch the auth metadata rollout. As servers adopt structured auth declarations, MCP clients will be able to auto-negotiate permissions. If you are building an MCP server, adding auth metadata now is worth the effort.

The registry will keep growing. The low barrier to entry is a feature, not a bug. Volume drives network effects, and quality filtering happens on top — not at the gate.

FAQ

Q: Is the official MCP registry run by Anthropic? A: Not directly. Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation in late 2025. The registry is a community-driven project under that foundation, with open-source code and open submissions.

Q: How many MCP servers are in the official registry? A: Approximately 9,652 as of late May 2026. The number is growing weekly. The broader ecosystem (including unregistered servers on GitHub and npm) exceeds 48,000.

Q: What is the difference between the official registry and a curated directory like AgentNDX? A: The official registry lists any spec-compliant server. Curated directories like AgentNDX add quality scoring, verification, install testing, and use-case tagging. The registry answers “what exists.” A curated directory answers “what works.”