AI safety and research company Anthropic has released the fifth specification of its Model Context Protocol, designated MCP 2026-07-28, marking the most substantial revision of the open standard since its introduction. The update arrives as MCP exceeds 400 million monthly SDK downloads, a fourfold increase this year, and strengthens its position as the dominant protocol for connecting AI agents to external applications, tools, and enterprise data sources.

The release fundamentally rearchitects MCP around a stateless core, replacing the prior bidirectional, session-dependent model with a lightweight request/response framework. By removing the need for servers to maintain persistent session state, the change enables deployment on serverless and edge infrastructure and allows straightforward horizontal scaling behind standard load balancers. That shift matters for teams that need connectors to run reliably in production environments where operational simplicity, uptime, and deployment flexibility are as important as raw capability. Anthropic said the reduction in operational overhead will let developers manage complexity more efficiently as connector usage grows.

Alongside the architectural change, the specification introduces a formal, versioned extensions framework designed to add capabilities without changing the protocol’s core. Three extensions graduate in this release: MCP Apps, which let servers render interactive user interfaces within sandboxed iframes directly inside AI conversations; Tasks, which support long-running and asynchronous operations beyond synchronous request boundaries; and Enterprise Managed Auth, which allows organizations to centralize access control through existing identity providers such as Microsoft Entra or Okta. Authorization mechanisms have also been hardened to align with production-grade OAuth 2.0 and OpenID Connect deployments, removing earlier enterprise integration workarounds. A formal deprecation policy has been established as well, giving developers a clearer path for adopting newer capabilities without breaking existing implementations.

Before, running a remote MCP server meant managing session state, which limited where you could run it. Now that MCP is stateless, you can deploy on serverless and edge infrastructure, or scale horizontally behind any load balancer. pic.twitter.com/2Rfs2kdgb8 — ClaudeDevs (@ClaudeDevs) July 28, 2026

Before, running a remote MCP server meant managing session state, which limited where you could run it. Now that MCP is stateless, you can deploy on serverless and edge infrastructure, or scale horizontally behind any load balancer. pic.twitter.com/2Rfs2kdgb8

Enterprise backing from Zoom, Figma, Intuit, and Netlify highlights production expansion

The update has drawn immediate endorsements from major ecosystem partners, underscoring its enterprise readiness. Zoom said stateless deployment improves reliability for bringing meeting intelligence into AI workflows, while Figma said horizontal scalability is critical for growing usage of its design-canvas connector. Intuit said the extensions framework supports agentic financial experiences at enterprise scale for its 100 million customers, and Netlify said treating MCP as a first-class HTTP workload removes session-management friction for platform developers.

Within Anthropic’s Claude platform, the protocol now supports more than 950 listed connectors used by millions of people daily. Recent product enhancements include inline interactive interfaces through MCP Apps, zero-touch enterprise provisioning via identity-provider group inheritance, an observability dashboard that lets developers track adoption and diagnose latency, and a research-preview feature called MCP tunnels that connects Claude to private-network servers without public endpoints or inbound firewall rules. Those additions point to the practical direction of the protocol: making it easier for enterprises to connect internal systems while keeping security controls and operational monitoring in place.

By moving to a stateless foundation, standardizing extensibility, and tightening enterprise authentication, MCP 2026-07-28 signals a deliberate move from an experimental integration layer toward production-grade infrastructure intended to support the broader agentic AI ecosystem.