In November 2024, Anthropic published a draft specification called Model Context Protocol. The pitch was practical: instead of every AI agent team writing custom integrations for every tool, why not a standard interface — a USB-C connector for AI agents to talk to data sources, APIs, and services?
Sixteen months later, MCP has crossed 97 million monthly SDK downloads, up from roughly 2 million at launch. That’s 4,750% growth. More importantly, it’s no longer Anthropic’s spec. It belongs to the industry — and OpenAI’s adoption has made it effectively permanent.
The Number That Actually Matters#
97 million downloads is a headline. The more important data point happened quietly alongside it: OpenAI added native MCP support to ChatGPT and simultaneously announced the deprecation of their Assistants API, with a sunset date in mid-2026.
That’s not adoption. That’s capitulation. When the company that invented the modern AI API model tells its developer ecosystem to migrate to someone else’s standard, that standard has won the coordination game. You don’t sunset your own API unless you’ve concluded that fighting a competitor’s standard isn’t worth the fracture.
This is what TCP/IP moments look like in practice. Not a triumphant announcement, but a quiet acknowledgment by competing parties that there’s more value in interoperability than in winning a standards war.
What 5,800 Servers Actually Mean#
The MCP ecosystem has grown to 5,800+ community and enterprise servers. The top 50 alone generate 170,000 monthly US searches. But the more instructive data comes from companies that have already standardized on MCP at scale:
- Block eliminated 340 custom connectors by migrating to MCP. Those aren’t minor internal utilities — each connector represents an integration someone built, maintained, and had to update when upstream APIs changed. MCP retired that entire engineering backlog.
- Apollo cut integration maintenance overhead by 60%. Integration maintenance is the invisible tax on every engineering team — the work that doesn’t ship features but breaks if you stop doing it.
- Replit built its entire AI development environment on MCP primitives. For a company where “AI runs the dev environment” is the product pitch, that’s not a tooling choice; it’s a foundational architecture decision.
These numbers explain the adoption curve. MCP crossed 97 million downloads not because developers love protocols, but because it made a concrete, measurable problem — the integration maintenance problem — significantly cheaper to solve.
From Anthropic’s Spec to Linux Foundation Governance#
In December 2025, MCP was donated to the Agentic AI Foundation (AAIF) under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI. This is the right move, done at the right time.
A protocol that lives inside one company’s GitHub repository won’t become industry infrastructure. Enterprises don’t standardize on something they can’t audit, fork, or hold someone accountable for. The Linux Foundation donation was the moment MCP stopped being Anthropic’s smart idea and became a public good with governance structure to match.
The analogy to earlier infrastructure standards is deliberate. HTTP lived inside CERN’s servers before Tim Berners-Lee donated it to the public domain. TCP/IP was a DARPA project before it became the Internet. The pattern is consistent: a technical standard becomes infrastructure when the inventor gives up control.
The Security Problem Everyone Saw Coming#
Adoption at scale surfaces the rough edges. In March 2026, Qualys published a report flagging MCP servers as “the new Shadow IT”: enterprise security teams discovering MCP deployments they didn’t authorize, authentication tied to static secrets, no standardized audit trails, and gateway behavior that varies by implementation.
Every MCP server is a potential attack surface. A rogue MCP server can request permissions an agent shouldn’t have, exfiltrate data through tool call responses, or inject malicious context into agent reasoning. The fact that MCP is agent-native makes these risks different from traditional API security — agents that can act on behalf of users are a larger blast radius than read-only API calls.
The 2026 MCP roadmap addresses this with async operations and hierarchical multi-agent support, but security standardization is the hard problem. The tools ecosystem needs a layer that handles authentication, authorization, and audit logging consistently across MCP implementations — or enterprises will build it themselves and create another fragmentation problem.
This is the predictable growing pain of any infrastructure standard. It’s not a reason to avoid MCP; it’s a reason to build security tooling around it, which is exactly what’s happening.
Why Claude Code Is the Primary Beneficiary#
Claude Code is a first-class MCP host. This wasn’t an accident — Anthropic designed MCP while building the agent systems that would use it. The integration is structural, not bolted on.
Every MCP server added to the ecosystem — database connectors, git tools, monitoring integrations, enterprise SaaS APIs — becomes immediately available to Claude Code agents without any code changes or custom integration work. When Block standardized on MCP, every Claude Code user working in that ecosystem got those 340 connectors for free.
This is a compounding flywheel:
- More MCP servers → more capable Claude Code agents
- More capable Claude Code agents → more developers building on Claude Code
- More Claude Code adoption → more incentive to build MCP servers
Cursor has MCP support. VS Code Copilot has MCP support. But Claude Code’s terminal-native architecture means the agent runs your tools rather than an IDE plugin suggesting how you might run them yourself. The MCP integration is load-bearing; in IDE-centric tools it’s a feature.
The Infrastructure Moment#
97 million downloads is the right headline because it signals something specific: MCP has crossed the threshold where it’s cheaper to build on the standard than to build around it. When Block retires 340 connectors, they’re not making an ideological statement. They’re making an engineering economics decision.
The protocol won because enough parties — Anthropic, OpenAI, Google, Microsoft, Block, Replit — decided that the coordination value of a shared standard outweighed the competitive value of proprietary connectors. OpenAI’s Assistants API sunset date formalized that decision.
For developers building on Claude Code: the MCP ecosystem is the toolbox. The question isn’t whether to use it — it’s which of the 5,800+ servers your agents should have access to.
Sources: