Origins and Development of the Model Context Protocol: From Internal Frustration to Public Launch

MCP (What Is the Model Context Protocol (MCP): the Open Standard Connecting AI Models to Tools and Data) started as one engineer's personal workflow annoyance inside Anthropic in mid-2024, moved through a six-week internal build and a hackathon, and became a public open standard on November 25, 2024. This page traces that arc: who built it, why, and what design choices were locked in at launch.

David Soria Parra's Frustration and the M×N Problem (July 2024)

David Soria Parra was a software engineer at Anthropic working on internal developer tooling (https://www.latent.space/p/mcp). Around July 2024 (https://newsletter.pragmaticengineer.com/p/mcp), he kept hitting a workflow wall: he constantly copy-pasted code between Claude Desktop and his IDE while building developer tools, because Claude Desktop had capabilities — like artifacts — that his IDE couldn't access (https://kylestratis.medium.com/the-surprising-origins-of-the-model-context-protocol-868d640ac7c6, https://www.latent.space/p/mcp).

He framed this as an "M×N problem": if every AI application needs a custom integration for every data source, the number of integrations to build grows multiplicatively. A concrete illustration from the sources: ten AI tools each needing to connect to a hundred data sources implies up to 1,000 custom integrations (https://kylestratis.medium.com/the-surprising-origins-of-the-model-context-protocol-868d640ac7c6). Solving this combinatorial-integration problem is the core motivation behind What Is the Model Context Protocol (MCP): the Open Standard Connecting AI Models to Tools and Data.

From Prototype to Product: Justin Spahr-Summers and Six Weeks of Building

David brought the concept to Justin Spahr-Summers, another Anthropic software engineer who brought product and infrastructure expertise. Justin "got excited pretty quickly and we started working on it" (https://www.latent.space/p/mcp).

The two worked together to build the first working MCP integration for Claude Desktop. Reported timelines vary slightly across sources: one source states "six weeks later had the first working MCP integration" (https://kylestratis.medium.com/the-surprising-origins-of-the-model-context-protocol-868d640ac7c6), while another (unverified against a second source) reports "roughly a month and a half building the protocol" (https://www.latent.space/p/mcp). Both point to a build time on the order of one to one-and-a-half months.

Internal Hackathon and Early Demos

Roughly one month before the public launch — approximately October 2024 — Anthropic held an internal hackathon (https://www.latent.space/p/mcp). Multiple teams rapidly adopted MCP and began building tools and integrations for Claude Desktop (https://kylestratis.medium.com/the-surprising-origins-of-the-model-context-protocol-868d640ac7c6).

Notable demo

one team built an MCP server that controlled a 3D printer, used as a real-world showcase of the protocol's reach beyond typical developer-tooling use cases (https://www.latent.space/p/mcp). The hackathon is described as having generated organic enthusiasm and confirmed MCP's real-world utility (https://newsletter.pragmaticengineer.com/p/mcp).

Design Philosophy at Launch

The launched protocol deliberately split functionality into three primitives: Tools (model-initiated function calls), Resources (data/context selected by the application or user), and Prompts (user-initiated text macros) — see MCP Core Primitives: Tools, Resources, and Prompts — who is allowed to invoke what, and what a server must declare to offer it (https://www.latent.space/p/mcp).

For the wire-level architecture, the team borrowed from the Language Server Protocol (LSP), adapting its message-flow model for AI integration rather than inventing a new one from scratch (https://en.wikipedia.org/wiki/Model_Context_Protocol) — see MCP Architecture: Host, Client, Server, and the JSON-RPC 2.0 Wire Format. On infrastructure, they deliberately stayed "boring": JSON-RPC as the transport, so that design effort could go into the primitives and features instead (https://www.latent.space/p/mcp). This pragmatism extended to ecosystem strategy: the team encouraged developers to "just start" building servers and let the ecosystem converge naturally, rather than over-specifying upfront (https://www.latent.space/p/mcp).

Unverified

the protocol is sometimes described publicly with a "USB-C for AI" analogy, but the notes found no explicit USB-C analogy in the primary sources checked (Anthropic announcement, Latent Space interview, Wikipedia). The underlying ideas of composability and standardization are confirmed; the specific analogy is not.

Public Open-Source Release on November 25, 2024

Anthropic officially announced and open-sourced MCP on November 25, 2024 (https://www.anthropic.com/news/model-context-protocol). The launch bundled the protocol specification, SDKs in programming languages including Python, TypeScript, C#, and Java, local server support in Claude Desktop, and an open-source repository of ready-made servers (https://en.wikipedia.org/wiki/Model_Context_Protocol, https://www.anthropic.com/news/model-context-protocol). Pre-built connectors shipped for Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer (https://newsletter.pragmaticengineer.com/p/mcp).

Anthropic framed the release around a commitment to building MCP as a "collaborative, open-source project and ecosystem" (https://www.anthropic.com/news/model-context-protocol) — a stance that set up its later transition described in From Anthropic Project to Linux Foundation Standard: How MCP Left Vendor Control. Framed against the original problem, MCP's promise was to turn the M×N integration burden into an M+N one: instead of 1,000 bespoke integrations for 10 tools and 100 data sources, each side only needs to implement MCP once (https://newsletter.pragmaticengineer.com/p/mcp).

What is missing from these notes

The notes do not cover: exact org-chart placement of David Soria Parra and Justin Spahr-Summers before MCP, internal approval/decision process at Anthropic, or any rejected alternative designs considered before settling on the three-primitive model. These gaps should be closed before raising this page's status to verified.