MCP ecosystem tops 13,000 servers as monthly SDK downloads hit 97 million

The ecosystem around the Model Context Protocol (MCP) has crossed 13,000 servers, and its SDKs are now downloaded roughly 97 million times a month, according to an ecosystem analysis published on dev.to. The numbers point to a protocol that is quietly becoming the default plumbing between AI assistants and the software those assistants are asked to operate.

The growth in numbers

The dev.to analysis, written by GrahamduesCN, puts the server count at more than 13,000 across npm and GitHub as of May 2026. Monthly SDK downloads have reached 97 million, about triple the level of six months earlier, and new server registrations are up 400 percent year over year. Anthropic's official reference servers still carry meaningful weight in that traffic: the filesystem server alone draws around 48,500 downloads per month, the post reports.

What MCP actually standardizes

A second dev.to post, an explainer by judaaaa, frames the problem the protocol addresses. A language model can parse an instruction such as creating a task in a project management app without having any way to act on it: comprehending a request and being able to carry it out are distinct capabilities, much as a frontend understands a user's intent but still needs a backend to do the work.

MCP defines a shared interface between AI applications and external systems such as calendars, task managers, databases and file systems. The explainer likens it to USB: rather than every device shipping its own connector, a common standard lets compatible things talk. One caveat the post stresses is that MCP does not replace the APIs behind a service. A calendar integration still communicates with the calendar provider; what the protocol standardizes is how an AI application discovers and invokes that integration, so the same server can be reused across MCP-compatible assistants such as ChatGPT and Claude.

Hosts, clients and servers

The architecture has three parts. The host is the AI application the user talks to, responsible for coordinating connections and deciding how external capabilities fit into a conversation. The client maintains a connection to an individual MCP server, handling protocol-level work such as discovering available tools. The server exposes capabilities, for example functions for creating tasks or querying a database, and talks to the underlying API or system to get the work done.

In a typical exchange, the user makes a request, the model decides a particular tool would help, the host coordinates the call through its MCP client, and the server performs the operation against the external service before the result flows back to the assistant. The explainer stresses that the model never calls a service like Todoist directly; it participates in a workflow the host orchestrates. That distinction matters, because a model deciding to request an action is not the same as that action completing successfully.

More than tool calls

Alongside tools, which perform operations, the protocol defines resources for exposing readable information such as configuration, documentation and database schemas, and prompts as a third core capability. The explainer also notes unfinished business: authentication, permissions, hosting and platform compatibility all still have to be handled outside the protocol itself.

The discovery gap

Rapid growth has produced a bottleneck of its own. According to the ecosystem analysis, locating the right server remains awkward, typically requiring improvised npm searches or trawling GitHub repositories. The author's response is mcp-hub, a command-line tool that lets developers search a registry and install servers directly, seeded with a handful of official packages that exist on npm. Planned next steps include private registries for enterprise teams, community submissions and CI/CD integration for auto-publishing servers.

Why it matters

The adoption curve suggests MCP is settling in as the connective tissue for agentic AI: build an integration once and any MCP-compatible assistant can use it, rather than rebuilding adapters for every platform. But an ecosystem of 13,000 servers also raises questions the protocol does not answer by itself, including discovery, quality, security and permissions. On this evidence, the gap between how quickly servers are being published and how hard they remain to find and vet is the next problem the community has to solve.