Model Context Protocol（MCP）发展史

History of the Model Context Protocol (MCP) Origins (mid-2024) MCP started as an internal fix for an annoyance. In July 2024, Anthropic engineer David Soria Parra got tired of manually copying answers from Claude into his code editor because the model had no way to reach outside tools. He and colleague Justin Spahr-Summers sketched out a protocol to solve that, along with Mahesh Murag. Public Launch — November 25, 2024 Anthropic introduced MCP as an open standard for connecting AI assistants to data systems like content repositories, business tools, and dev environments, shipping it with SDKs for Python and TypeScript. The goal was fixing the "N×M" integration problem — needing a custom connector for every AI-tool pairing. 2025 — Rapid Iteration and Adoption March 2025: Anthropic shipped a major spec update adding OAuth 2.1 for secure agent-server communication and Streamable HTTP Transport for real-time bidirectional data flow. Around this time thousands of community MCP servers began appearing on GitHub. Through the year, MCP was adopted well beyond Anthropic — OpenAI, Google DeepMind, Microsoft, and countless developers building production agents all took it up. November 2025: Another major spec update added asynchronous operations, statelessness, server identity, and an official community-driven registry for discovering MCP servers. Handoff to Open Governance — December 2025 Anthropic donated MCP to the newly formed Agentic AI Foundation (AAIF) under the Linux Foundation, making it a vendor-neutral standard rather than an Anthropic-owned project. 2026 — Mainstream Infrastructure By mid-2026, MCP had become genuinely ubiquitous: ChatGPT, Claude, Gemini, Microsoft Copilot, VS Code, and Cursor all speak it, and its two largest SDKs pulled more than 470 million downloads in the 30 days ending July 21, 2026. Other sources from earlier in the year cite figures like 97M+ monthly downloads and 5,800+ servers, showing just how fast adoption climbed over just a few months. Related: Skills There's also a related but separate thread worth knowing about: Anthropic has more recently been pushing Skills (lightweight, on-demand instruction sets) as a complementary approach — since MCP connections can be resource-intensive while Skills are lighter-weight and preserve context better, though the two are described as overlapping rather than competing.