The MCP Protocol Fundamentals article covers the full protocol in one pass. This hub organizes the underlying concepts into standalone garden notes for reference and deeper reading.
If you’re new to MCP, start with MCP Protocol Architecture and work through the notes in order. If you need a specific reference, jump directly to the relevant note.
The Foundation
MCP Protocol Architecture — What MCP is, why it exists (the N×M problem, the USB-C analogy), the three-layer architecture (host/client/server), all three server primitives (tools, resources, prompts), the security model, and how MCP compares to traditional REST APIs. The starting point for everything else.
MCP Ecosystem Governance — The adoption story and governance trajectory: Linux Foundation donation, OpenAI and Microsoft adoption, the 97M SDK download milestone. If you’re deciding whether to invest in MCP, this note answers the infrastructure durability question.
Protocol Internals
MCP JSON-RPC Wire Format — The actual messages on the wire: initialization handshake, capability negotiation, tool discovery (tools/list), tool invocation (tools/call), and error handling. Essential for debugging production issues and understanding what the SDK is doing for you.
MCP Client Primitives — The bidirectional side of MCP: sampling (server-requested LLM completions), elicitation (server-requested user input), and roots (filesystem boundaries). Most servers don’t use all three, but understanding them unlocks multi-step reasoning workflows and principled security boundaries.
Primitives in Depth
MCP Resources and Prompts — Goes deeper on resources (read-only data exposure via URI templates, when to use vs. tools) and prompts (reusable templates for standardized workflows). Includes the Context object for progress reporting on long-running operations.
MCP Transport Configuration — Practical setup for both transports: stdio for local development (zero-config, credential isolation) and streamable HTTP for production (multi-client, OAuth 2.1, load balancing). When to switch and how.
Building Servers
FastMCP Server Skeleton — The minimal working server in Python and TypeScript. The starting point for any custom server.
MCP Tool Design Patterns — How to write tools the AI uses correctly: docstrings as descriptions, Pydantic for structured output, input validation. The difference between tools that work and tools the AI fumbles.
MCP Server Testing and Debugging — The Inspector, the stdout-corrupts-protocol gotcha, and the testing workflow.
Related
MCP Ecosystem Overview — The full landscape: official servers, community servers, clients, and where to find things.
Security Posture for AI Agents — Security principles for AI tools accessing data infrastructure. Relevant for any production MCP deployment.
CLI vs MCP for AI Agents — When native CLI commands are more efficient than MCP. Token efficiency, training data advantages, and the cases where MCP wins.