MCP Won: The USB Moment for AI
The Agent Era — Series
Episode 7 of 10
The Agent Era — Series
Episode 7 of 10 · ← Previous · Next →
The Universal Plug
In the early days of personal computing, every device had its own cable. Printers, scanners, keyboards, mice — each required a proprietary connector. Then USB arrived, and within a decade, the chaos became order. One standard. One port. Everything just worked.
AI is experiencing its USB moment right now. The standard is called MCP — the Model Context Protocol — and its adoption curve suggests it has already won.
The Numbers Are Extraordinary
- 97 million+ monthly SDK downloads — outstripping even React's journey to 100 million.
- 10,000+ public MCP servers available for developers.
- 13,000+ GitHub repositories containing MCP server implementations.
- 28% of Fortune 500 companies have adopted MCP in some capacity.
These are not early-adopter numbers. This is mainstream infrastructure.
A Brief History
MCP began as an Anthropic project — a protocol designed to give Claude a standardized way to connect to external tools and data. In December 2025, Anthropic made a pivotal decision: they donated MCP to the Linux Foundation, establishing the AI Agent Foundation (AAIF) as its governing body.
The founding co-sponsors included Anthropic, Block, and OpenAI — a rare moment of cooperation among fierce competitors. The platinum sponsors: AWS, Google, Microsoft, Cloudflare, and Bloomberg. When Amazon, Google, and Microsoft all agree on a standard, the debate is over. MCP won.
Platform Support
- Claude ships with 75+ built-in MCP connectors.
- ChatGPT and Copilot have added native MCP support.
- Cursor ($500M ARR, 360K paying users) uses MCP as its primary extension mechanism.
- Google Gemini and Bloomberg Terminal have integrated MCP servers.
According to BCG, MCP reduces integration complexity from quadratic to linear. Before MCP, connecting N models to M tools required N×M custom integrations. With MCP, it requires N+M.
The Broader Protocol Landscape
- MCP — the tool interface layer (how agents connect to tools and data).
- A2A (Agent-to-Agent, from Google) — the inter-agent communication layer.
- AGENTS.md (from OpenAI) — the coding guidance layer.
Together, they form a complete stack: context, communication, and execution.
The Security Problem
- 30+ CVEs disclosed in January-February 2026 alone.
- 43% of tested MCP servers vulnerable to command injection.
- CVE-2025-6514 affected an estimated 437,000 environments.
Security researchers have drawn a direct parallel: “This is Log4j all over again.” A ubiquitous, deeply embedded protocol with inadequate security guardrails is exactly the kind of systemic risk that keeps CISOs awake at night.
The protocol war is over. MCP won. But the real battle — building agents that actually work reliably — has just begun.
MCP solves the connectivity problem. But a protocol is not an architecture. The framework you choose to build on top of MCP determines whether your agents are production-ready or just another failed pilot.
Comments ()