David Soria Parra from Anthropic outlines the current state and future direction of the Model Context Protocol (MCP). MCP has reached 110 million monthly downloads in roughly half the time React did, with adoption across OpenAI, Google ADK, LangChain, and thousands of frameworks. Key themes include: MCP applications that let agents ship their own UI interfaces; the shift from coding agents to general knowledge-worker agents in 2026; a 'connectivity stack' combining skills, MCP, and CLI/computer use; and two critical client-side patterns — progressive discovery (lazy-loading tools on demand to reduce context bloat) and programmatic tool calling (having the model write scripts rather than chaining individual tool calls). On the protocol roadmap: a stateless transport protocol from Google landing in June, improved async task primitives for agent-to-agent communication, TypeScript and Python SDK v2, cross-app SSO access for enterprises, server discovery via well-known URLs, and skills-over-MCP for shipping domain knowledge alongside servers.
Sort: