MCP Hits 97 Million Downloads: The Protocol That Won

Anthropic's Model Context Protocol just crossed 97 million monthly SDK downloads. Seventeen months ago it didn't exist. The protocol launched in November 2024 as an experimental spec for connecting AI apps to external tools. By March 2026 it has 10,000+ active public servers, every major AI provider ships MCP-compatible tooling, and it has quietly become the universal standard for how AI agents talk to anything that isn't the model itself. This is the rarest kind of protocol story: a single company proposed a spec, and within two years the entire industry shipped it.

If you're not paying attention to MCP yet, you're about to be forced to. It is to AI agents what LSP was to code editors and USB was to hardware — the integration layer that makes everything else downstream work.

What MCP Actually Is

MCP is a spec. Specifically, a JSON-RPC-based protocol that defines how a client (an AI application) asks a server (a tool or data source) what it can do, and then how to do it. You can think of it as a standardized vocabulary for AI agents to discover tools, call them, and interpret responses.

Before MCP, if you wanted to give an AI assistant access to your database, your issue tracker, and your file system, you wrote three custom integrations — one for each combination of AI client and tool. N clients times M tools equals a lot of glue code. With MCP, you write one server per tool, and every MCP-compliant client gets it for free.

This is the same pattern that made LSP (Language Server Protocol) successful. Before LSP, every editor wrote its own integrations for every language. After LSP, you wrote one Go language server and it worked in VS Code, Neovim, Sublime, and Emacs. MCP is doing the same thing for AI agents — one server, many clients.

How We Got to 97 Million

Protocols usually die slow, public deaths. MCP didn't. The growth curve is worth looking at because it tells you when each major adopter joined and how fast each one pulled the rest of the industry along.

The pattern is textbook network effect. Anthropic launches at 2M. OpenAI joins in April 2025 and downloads jump to 22M. Microsoft integrates it into Copilot Studio in July 2025 — 45M. AWS adds support in November 2025 — 68M. By March 2026 it's sitting at 97M and still climbing.

What to Build With It

Practical use of MCP is simpler than the spec suggests. If you're a builder, you have two things to decide. Do you want your AI app to consume existing MCP servers, or do you want to expose your own tools as an MCP server for others to consume? Usually both, eventually.

from mcp.server import Server
from mcp.types import Tool

app = Server("internal-crm")

# Declare a tool any MCP-compatible client can call
@app.tool()
async def lookup_customer(email: str) -> dict:
    """Look up a customer by email in the internal CRM."""
    row = await db.query(
        "SELECT id, name, plan, mrr FROM customers WHERE email = $1",
        email,
    )
    return {"customer": row, "found": row is not None}

if __name__ == "__main__":
    app.run(transport="stdio")

That's the entire server. Run that file, add its path to your Claude Desktop or Cursor config, and your AI assistant can now look up customers in your CRM. No custom plugin. No vendor lock-in. No glue code per client. This is the whole pitch.

What It Means for Builders

The Bottom Line

If your AI app still has custom integrations for every tool it connects to, you're doing duplicate work the rest of the industry stopped doing six months ago. Pick one tool in your stack this week, wrap it as an MCP server, and watch the complexity of your integration code collapse.