David Soria Parra's Model Context Protocol

Table of content

David Soria Parra co-created the Model Context Protocol while working at Anthropic, starting from a simple annoyance: constantly switching between Claude Desktop and his code editor to copy-paste context. That frustration became an open standard now supported by ChatGPT, Cursor, Gemini, and VS Code, with over 97 million monthly SDK downloads.

Background

Member of Technical Staff at Anthropic (joined April 2024)
Co-creator and Lead Core Maintainer of MCP
Previously at Meta/Facebook (2013-2023), leading 45+ person org focused on static analysis and developer tools
Entrepreneur in Residence at Sutter Hill Ventures (2023)
Core PHP developer who shipped PHP 5.4 and 5.5 releases
Built Mercurial features: bookmarks, pager, shelving
Education: Karlsruhe Institute of Technology (2007-2013)
Based in London

GitHub | Twitter | Site | MCP Docs

MCP Origins

The protocol started at an Anthropic hackathon in 2024. Soria Parra and Justin Spahr-Summers wanted to solve what he calls “the M times N problem”: multiple AI clients needing to connect to multiple data providers. Without a standard, every integration requires custom work.

“MCP tries to enable building AI applications in such a way that they can be extended by everyone.”

Think USB-C: one connector instead of a different cable for every device. MCP provides:

Component	Purpose
Tools	Functions the AI can call (file operations, API requests)
Resources	Data the AI can read (files, database schemas)
Prompts	Reusable prompt templates servers can expose
Sampling	Servers can request LLM completions back from the client

Architecture

MCP uses a client-server model where the AI host connects to multiple servers, each exposing capabilities:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/files"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "your-token"
      }
    }
  }
}

The protocol supports stdio (local processes) and HTTP with Server-Sent Events (remote servers). Claude Desktop, Cursor, and VS Code all support MCP as clients.

M x N Problem

Before MCP:

5 AI clients x 10 data sources = 50 custom integrations

With MCP:

5 AI clients implement MCP client once
10 data sources implement MCP server once
Everything connects

This is why adoption spread fast. OpenAI, Microsoft, Google, Docker, and GitHub all added support. Building one MCP server makes your tool accessible to every MCP-supporting AI client.

Underused Features

Soria Parra notes that most developers only use tools, ignoring other primitives:

Feature	What It Does
Sampling	Server requests LLM completions from the client (multi-agent patterns)
Resources	Expose readable data with URIs and MIME types
Prompts	Server-provided prompt templates users can invoke
Notifications	Real-time updates between client and server

Sampling is the interesting one. An MCP server can request LLM completions from the client, so your server could analyze data, ask Claude for a summary, then return structured results. Multi-agent patterns without managing API keys on the server side.

Open Source Governance

In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. Co-founded with Block and OpenAI, the foundation also hosts goose (Block’s AI agent) and AGENTS.md (OpenAI’s agent configuration standard).

Soria Parra remains Lead Core Maintainer. The donation means:

Neutral governance
No single company controls the spec
Infrastructure support from Linux Foundation
Same maintainers making technical decisions

Key Takeaways

Principle	Implementation
Solve your own problem	Copy-paste frustration became a protocol
Protocol over products	M x N reduction from standardization
Open source from day one	Community contributions shaped the spec
Underutilized features matter	Sampling, resources, prompts need documentation

Jason McGhee’s WebMCP - Turning websites into MCP servers
Three-Layer Workflow - Spec, implement, verify patterns
Context Management - Managing what AI can see

Links

Next: Simon Willison’s LLM Workflow