How MCP grew from an internal Anthropic hack into AI's standard tool protocol

An engineer's workaround, not a moonshot

According to a retrospective on dev.to, the Model Context Protocol began in July 2024 as a fix for a personal annoyance rather than a strategic initiative. David Soria Parra, an engineer at Anthropic, was tired of shuttling Claude's answers into his code editor by hand because the model had no way to reach external tools. Together with colleagues Justin Spahr-Summers and Mahesh Murag, he sketched out a protocol to close that gap.

Launch as an open standard

Four months later, on November 25, 2024, Anthropic published MCP as an open standard for connecting AI assistants to data systems such as content repositories, business software and development environments, the post recounts. It shipped with SDKs for Python and TypeScript. The motivating problem was combinatorial: without a shared protocol, every model needed a bespoke connector for every tool, an overhead the post labels the N times M problem. One standard, the thinking went, collapses that matrix into a single integration surface.

A fast year of spec changes and adoption

The 2025 timeline in the dev.to account shows unusually rapid iteration for a young specification. A March 2025 update added OAuth 2.1 for securing agent-to-server communication and a streamable HTTP transport for real-time, bidirectional data flow. Around the same time, community-built MCP servers began appearing on GitHub by the thousands.

Adoption also spread well beyond Anthropic. OpenAI, Google DeepMind and Microsoft, along with developers running production agents, took up the protocol during the year. A second major update in November 2025 brought asynchronous operations, a stateless mode, server identity and a community-run registry for discovering servers.

Handoff to neutral governance

In December 2025, according to the post, Anthropic donated MCP to the newly formed Agentic AI Foundation under the Linux Foundation, moving the protocol out of single-company ownership and into neutral stewardship. That transfer matters for anyone betting an integration roadmap on MCP, since its future direction no longer sits inside any one lab's product priorities.

Ubiquitous infrastructure by mid-2026

By the middle of 2026, the post claims, MCP had become genuinely pervasive: ChatGPT, Claude, Gemini, Microsoft Copilot, VS Code and Cursor all support it. The two largest SDKs logged more than 470 million downloads in the 30 days ending July 21, 2026. Earlier figures the post cites from other sources, such as more than 97 million monthly downloads and over 5,800 servers, show how steep the curve already was months before that. The spread between those numbers and the mid-year figure also illustrates how sharply the ecosystem compounded over a single year.

A parallel track: Skills

The dev.to piece closes with a related thread worth tracking. Anthropic has more recently been promoting Skills, small on-demand bundles of instructions, as a complementary approach to MCP. Connections over MCP can be comparatively resource-heavy, while Skills are lighter and gentler on context. The two are described as overlapping rather than competing, but the split suggests builders may end up choosing between heavy tool connections and light instruction formats depending on the task at hand.

Why it matters

MCP's two-year climb from an internal hack to a Linux Foundation-governed standard explains why the AI tooling landscape consolidated around one integration surface so quickly. For developers, it means writing a connector once and reaching every major assistant and editor. For the industry, it is a case study in how fast de facto standards can crystallize when a protocol is open, minimal and adopted early by rival vendors. The Skills development is the caveat: MCP may be the plumbing, but the division of labour between it and lighter instruction formats is still being worked out.