The Creators of Model Context Protocol — Anthropic's MCP Origin Story, Design, and Roadmap
In this Latent Space episode, Anthropic engineers Justin Spahr-Summers and David Soria Parra — the co-creators of MCP — give the first major podcast deep-dive on the Model Context Protocol, which they frame as the "USB-C port of AI applications" and the winner of the agent open-standard wars of 2023–2025. MCP is a client-server protocol that solves the M×N integration problem for AI apps, borrowing JSON-RPC and bidirectionality from LSP but deliberately diverging in key design choices. Anthropic positioned the protocol around application developers (not models directly), emphasizing that the client app and user remain in full control of what is exposed and invoked. The conversation digs into MCP's three core primitives — tools (model-initiated actions), resources (application- or user-controlled data/context, URI-addressed), and prompts (user-initiated templates/multi-step interactions) — and argues the ecosystem over-relies on tool calling. The hosts discuss MCP vs OpenAPI (complementary, not competing), the move from SSE to the new "Streamable HTTP" transport to enable horizontal scaling while preserving statefulness, draft OAuth 2.1 authorization, and registry/trust concerns (advocating the MCP Inspector over social-proof-based vetting). They also explore composability — recursive servers that are also clients, building agent graphs/DAGs out of MCP — while warning against protocol bloat ("don't try to do everything"). On governance, both guests stress MCP must remain an open, multi-company project (SDKs already maintained by Microsoft, JetBrains, Spring AI, Pedantic, etc.) but are wary of formal standards bodies slowing iteration. They highlight the AI Engineer World's Fair MCP track (June 3–5, 2025) and call for more full-spec clients (especially sampling clients), richer uses of resources and prompts, and integrations like Godot/Blender. Notable admissions: several reference servers were largely written by Claude itself, illustrating Anthropic's dogfooding of its own model.
Original sources
Podcast Latent Space
YouTube Latent Space
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.