Anthropic published the Model Context Protocol on 25 November 2024 with about as much fanfare as any open technical spec gets — a GitHub repo, a blog post, a handful of example servers. It had roughly 100,000 SDK downloads in its first month. Sixteen months later, that number is 97 million downloads a month. Every major AI lab — OpenAI, Google, Microsoft, Salesforce — shipped support for it within thirteen months of launch. Governance of the protocol now sits with the Linux Foundation, under its Agentic AI Foundation, which is the kind of neutral institutional home a technology only gets once it’s too load-bearing for one company to keep unilaterally.

That’s an unusually fast path from “one company’s spec” to “industry infrastructure,” and it happened because MCP solved a specific, boring, expensive problem that every AI lab was independently reinventing.

For the rest of us: the problem MCP actually solves

Before MCP, if you wanted an AI model to check your calendar, search a database, or read a file, someone had to write a custom integration — specific to that model, that tool, and usually that vendor’s particular way of describing what the tool does. Ten tools meant ten bespoke integrations. Ten AI applications wanting to use the same ten tools meant a hundred integrations, most of them doing nearly identical work in slightly incompatible ways.

Model Context Protocol is a standard connector. Think of it the way USB-C replaced a drawer full of proprietary chargers: instead of every device needing its own cable, one connector works everywhere, because both sides agreed on the same interface. An MCP server exposes a tool, a data source, or a set of prompts in one standard shape. Any MCP-compatible AI application — Claude, or increasingly, competitors’ tools too — can plug into it without custom integration work. Build the connector once; every compliant client can use it.

That’s the entire idea, and its simplicity is exactly why it spread as fast as it did — it removed a category of work that added no differentiated value for anyone building it.

What actually grew, and how fast

The ecosystem numbers are the clearest evidence that this went from novelty to default. There were roughly fifty MCP servers in existence at launch. By late May 2026, the official MCP Registry held 9,652 current server records — 28,959 total server-and-version records, counting every published update — and estimates across official and community directories put the real number of distinct servers somewhere between eight and twelve thousand by the middle of the year.

Adoption inside companies tells a more measured story than the download curve alone suggests. Nearly half of software companies report using MCP somewhere in their stack, but only 19% have it in broad production use, and across all MCP deployments industry-wide, just 41% make it to production at all. That gap — huge experimentation, much more modest production deployment — is normal for infrastructure this new, and it’s worth sitting with rather than smoothing over: a connector standard being everywhere in prototypes and comparatively rare in production is a maturity signal, not a failure signal.

The part that matters for anyone deploying this

Every MCP server an AI agent connects to is something that agent now implicitly trusts — it can describe tools in whatever way it wants, and by default, the agent takes that description at face value. That’s not a flaw unique to MCP; it’s the same trust boundary every plugin and extension ecosystem has always had to manage. But MCP’s speed of adoption means a lot of organizations are wiring agents into servers faster than they’re building the vetting process for what those servers are actually allowed to do once connected.

This is the same supply-chain question that showed up in the CrewAI/DSPy backdoor incident earlier this year, just from the connector angle instead of the package-registry angle: a standard that makes integration trivially easy also makes it trivially easy to integrate something you haven’t actually vetted.

What this means

MCP won the argument about whether AI tools needed a shared connector standard — the download curve and the lab adoption both say that argument is over. What’s still being worked out, in real time, across thousands of organizations right now, is the governance layer that should have shipped alongside it: who decides which servers an agent is allowed to trust, what a server is allowed to claim about itself, and how you audit that after the fact. The protocol solved the integration problem. It didn’t solve the trust problem, and that one doesn’t have a standard yet.

References