MCP: The Boring Spec That Became AI's USB-C
Almost two years ago, Anthropic open-sourced a message format — not a model, not a product, not a platform. The Model Context Protocol (MCP) landed on November 25, 2024, and the AI industry, which rarely agrees on anything, quietly made it the closest thing it has to a USB-C port.
MCP is deliberately unglamorous. An AI application (the client) connects to a server that exposes three kinds of things: tools it can call, resources it can read, and prompt templates it can follow. The transport is JSON-RPC 2.0 over Streamable HTTP, and the session starts with an initialize handshake where both sides advertise their capabilities. If that sounds like plumbing, that's the point. Before MCP, every pairing of a model and a tool meant bespoke glue code — the industry's N×M problem, where each new model-tool pair needed its own connector.
The protocol was built at Anthropic by David Soria Parra and Justin Spahr-Summers, and it launched with partners already on board: Block, Apollo, and the IDE crowd — Zed, Replit, Codeium, Sourcegraph. The pitch: write one server, and every compliant client gets it.
The next five months were the interesting part. On March 26, 2025, OpenAI adopted a rival's standard — Sam Altman, in the flat-footed phrasing of someone accepting the obvious, said "People love MCP" — shipping support in the Agents SDK that same day and joining MCP's steering committee in May. Google DeepMind's Demis Hassabis followed on April 9; Google Cloud shipped managed remote MCP servers for Maps, BigQuery, and Kubernetes in December. Microsoft went further at Build 2025, making MCP native to Windows 11, with a registry and OS-level servers exposing the file system and Windows Subsystem for Linux. It arrives switched off and auditable, which tells you how much trust a protocol this young commands.
The nicest detail comes from Microsoft's other team. VS Code's engineers — who invented the Language Server Protocol in 2016 — wrote in April 2025 that LSP's ideas inspired MCP. By June, VS Code supported the full specification, and version 1.102 made it generally available in July.
Then the numbers took off, for a spec. The official registry listed more than 10,000 active servers by December 2025, when Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation. A September 2026 count of the same registry found 35,286. The official Python and TypeScript SDKs were pulling about 97 million downloads a month within a year of launch; by mid-2026, roughly 16,000 GitHub repositories carried the mcp-server topic. GitHub's own official server went from public preview in April 2025 to generally available that September, with OAuth 2.1 and PKCE.
None of this means it's finished. Adoption has run ahead of security: MCP servers are code that executes with your agent's privileges, and prompt injection is the standing worry — GitHub's server now sanitizes content by default and offers a Lockdown mode for untrusted repositories. Discovery is still messy, and plenty of servers are one-afternoon experiments. The late-2025 specification added experimental asynchronous "MCP Tasks," and a paper on production deployments this year flagged missing primitives like identity propagation and structured error handling.
None of that has slowed it down, and that's the lesson. AI's loudest arguments are about models; its real progress lives in boring agreements. Write one MCP server and every compatible client — Claude, ChatGPT, Cursor, VS Code, Replit, or whatever wins next year — can use it. Two years in, the standard that changed the industry is a message format, and it was open-sourced by the underdog.