MCP origin story — solving the M×N integration problem between AI apps and external tools¶
Insight: MCP was created by two engineers (David and Justin) at Anthropic starting July 2024, born from frustration at constantly copying things between Claude Desktop and an IDE. It's designed to extend AI applications (not models themselves) — "the USB-C port of AI applications." The protocol solves the M×N problem: multiple AI apps need to integrate with multiple external tools, and a standardized protocol avoids each integration being custom. MCP is explicitly inspired by LSP (Language Server Protocol).
Detail: The timeline: David pitched the idea to Justin in July 2024 → a month and a half of building "unrewarding" groundwork (communication protocol, SDKs) → internal Anthropic hackathon energized the project (someone built an MCP server controlling a 3D printer) → first implementation was in Zed editor (open-source, written by David) a month before official release → public launch November 25, 2024. By April 2025, OpenAI and Google had adopted MCP support, confirming it as the winner of what Latent Space calls the "agent standard wars." MCP overtook OpenAPI in GitHub stars within months. The March 2025 spec update added stateless/resumable/streamable HTTP transports and OAuth 2.1 authorization.