Anthropic's Agent Standard Tops 400 Million Downloads a Month

Two years ago, connecting an AI model to a calendar, a database, or a company's internal tools meant writing a bespoke bridge for every pairing. Anthropic proposed a common plug instead, called it the Model Context Protocol, and gave it away. That bet has now paid off in a number worth pausing on: MCP passed 400 million monthly downloads of its software kits, about four times where it stood a year ago.

MCP is an open standard. It gives any model a single, predictable way to reach external tools, data, and services, so a developer writes one connector and it works across systems. Anthropic introduced it in November 2024, and the notable part is who lined up behind it. OpenAI, Google DeepMind, and Microsoft all adopted the same protocol, an unusual show of agreement in a field where the big labs rarely standardize on anything. In December 2025, Anthropic handed the project to the Agentic AI Foundation under the Linux Foundation, which turned it into a vendor-neutral standard rather than one company's house style.

The upgrade behind the surge

The download jump did not come from marketing. It tracks a genuine change in how MCP works. The 2026-07-28 specification reworked the protocol around a stateless core, moving it from a chatty, always-connected design to a plain request and response model. In practice that means MCP servers can now run on serverless and edge platforms like AWS Lambda, Vercel, and Cloudflare Workers, where they spin up on demand and cost nothing when idle. For teams deploying agents at scale, that is the difference between a science project and something they can put in production.

It fits a pattern we have watched build all year. Anthropic has been steadily turning MCP from a clever idea into plumbing that enterprises can trust, including handing IT admins control over which connectors employees can use. Google, meanwhile, has pushed a companion standard for agents talking to each other, the Agent2Agent protocol, now housed at the same foundation.

Why a boring number matters

Download counts are easy to wave away. This one is a decent proxy for something real, though. When a protocol becomes the default way an entire industry wires models to the outside world, whoever wrote it helps set the terms for how AI agents get built. Anthropic gave MCP away and still ended up holding that pen. For a company that also sells one of the frontier models plugging into all those tools, that is a comfortable place to stand.