MCP Just Went Stateless — and Anthropic's Numbers Say Adoption Already Outran the Old Architecture

Anthropic shipped MCP 2026-07-28, the fifth major revision to the Model Context Protocol, and the headline number explains why the architecture had to change: MCP has surpassed 400 million monthly SDK downloads, a 4x year-over-year increase, with Claude’s connector directory now listing more than 950 MCP servers. That’s the scale problem this update actually solves. The protocol moves from a bidirectional, stateful design to a stateless request/response core, letting servers run on serverless and edge infrastructure instead of requiring persistent connections. MCP Apps and Tasks move under a versioned extensions framework, giving developers a formal path to add interactive UIs and long-running work without touching the core spec, and authorization now aligns with production OAuth 2.0 and OIDC, plugging into enterprise identity systems like Entra and Okta. Figma, Intuit, Netlify, PostHog, Xero, and Zoom all contributed to the beta.

What enterprises are already building on top of MCP shows the stakes of getting this architecture right. Apollo’s CEO Matt Curl describes shipping a custom sequence-analytics tool in 30 minutes for roughly $30 in token costs — work that used to require a RevOps analyst and a data-warehousing budget — using the same native MCP integrations (Claude, Codex, Cursor, GitHub Copilot) this update formalizes. But Forrester’s David Mooter has been warning that most enterprises are building this capability the wrong way: his research argues an agent without proper integration governance is “merely an answer engine,” and that MCP servers need the same versioning, access control, and lifecycle management API teams already run, not a parallel, ungoverned stack. Mooter’s framing is blunt: organizations prioritizing integration and change management over the next year are the ones with real agentic-AI readiness; the ones chasing model capability alone aren’t.

Enterprise OAuth support is Anthropic’s answer to exactly that governance gap. Whether it’s sufficient on its own, or whether Mooter’s fuller integration-governance stack is still needed on top, is the question consulting engagements built around MCP adoption will spend the next two quarters answering.