MCP 2026-07-28: The Model Context Protocol Comes of Age – Stateless, Scalable, and Enterprise-Ready

July 2026 marks a turning point for the Model Context Protocol (MCP): the 2026-07-28 specification represents the biggest leap forward to date. The open-source protocol initiated by Anthropic, often called “the USB-C of AI tool integration,” has thus grown from an experimental standard into production-ready infrastructure. For anyone developing or operating AI agents, a closer look at the innovations is worthwhile.

From Session Mandate to Stateless Core

By far the most important change is the shift to a stateless Protocol Core. Where MCP previously required a complex initialize/initialized handshake and an Mcp-Session-Id, every request is now self-contained. Protocol version, client identity, and capabilities travel within the _meta field of each call. A client can query a server’s capabilities using the new server/discover method — but it doesn’t have to.

The practical impact is enormous: An MCP server that previously needed sticky sessions, a shared session store, and deep packet inspection at the gateway can now run behind a simple round-robin load balancer. Gateway operators can route requests using the new HTTP headers Mcp-Method and Mcp-Name without having to parse the JSON body. List results (tools/list, prompts/list) also carry cache hints (ttlMs, cacheScope), allowing clients to cache tool catalogs across multiple calls.

Those who still need state across multiple calls can manage it themselves via explicit handles (e.g., a basket_id) — a pattern that, according to the maintainers, is often even more powerful than the old session model because the state becomes visible and combinable for the model.

Extensions Become First Class

With the new specification, MCP introduces a formal Extensions Framework. Two extensions stand out in particular:

MCP Apps enable server-side rendered user interfaces. An MCP server can now deliver not just JSON data but interactive UI components — comparable to apps in an IDE extension. This opens doors for dashboards, configuration masks, and visual feedback loops directly within the agent workflow.

Tasks moves from the experimental core into its own extension and receives a poll-based tasks/get as well as a new tasks/update. Change notifications will in future run through a central subscriptions/listen stream, which clients subscribe to per notification type.

Multi Round-Trip Requests: The Server Asks Back

Previously, MCP required a permanently open SSE connection for queries back to the user — for instance, for confirmation before deleting files. With Multi Round-Trip Requests (MRTR, SEP-2322), this is no longer necessary. The server responds with resultType: "input_required" and lists the required inputs. The client collects the responses and repeats the original call with appended inputResponses. Any server instance can handle the repetition because all necessary information is contained in the payload. Supabase, for example, was precisely waiting for this function to get costly or destructive actions confirmed before execution.

Authorization: Enterprise-Ready

The security architecture has been hardened in several points. Authorization servers must now deliver the iss parameter according to RFC 9207, and clients must validate it — this closes a dangerous vulnerability where an attacker could substitute a foreign authorization server. Dynamic Client Registration (DCR) is being phased out in favor of Client Metadata Documents (CIMD). The transition runs over twelve months, but new implementations should directly use CIMD.

Roots, sampling, and logging are officially deprecated. They will still work for at least twelve months, but new builds should no longer use them. The legacy HTTP+SSE transport is also considered deprecated.

The SDKs Are Ready

All four Tier-1 SDKs (TypeScript, Python, Go, C#) support the 2026-07-28 specification, and the Rust SDK supports it in beta stage. The download numbers speak volumes: The MCP SDKs collectively register nearly half a billion downloads per month, with TypeScript and Python each having surpassed the billion mark in total downloads.

Conclusion

Over the last 18 months, MCP has evolved from a niche protocol into the dominant integration layer for AI agents — 78% of enterprise AI teams are already using it in production, and 67% of CTOs name MCP as their standard. The 2026-07-28 specification solves the central performance problem (state management) and simultaneously delivers enterprise features like hardened authorization, cacheable lists, and a formalized extensions model. Anyone starting with MCP or migrating today is investing in a platform that will form the backbone of agentic AI integration for years to come.

Sources

• MCP Blog: The 2026-07-28 Specification — Official release announcement with full changelog
• MCP Blog: 2026-07-28 Release Candidate — Technical deep dive into the RC, stateless core and MRTR
• RockB: MCP Enterprise Adoption Guide 2026 — Adoption statistics, gateway patterns, security framework
• DEV Community: MCP in 2026 – The USB-C of AI Tooling — Practical perspective on model-agnostic tooling
• MCP Blog: The 2026 MCP Roadmap — Priority areas and working group structure