What happened
The final MCP 2026-07-28 protocol specification is expected on July 28, 2026, and it is the largest update since the protocol launched. The headline change is the move to a stateless architecture: sessions and the initialize/initialized handshake are gone. Three core protocol features are being deprecated, and the authorization mechanism is fully rewritten to fit modern agentic scenarios. On top of that, a new extensions framework arrives, giving teams a flexible way to grow the protocol's functionality.
The release candidate (RC) of this spec shipped back on May 21, giving the community a ten-week window to validate and migrate. The point of that window is to let SDK and server developers adapt their products to the new requirements. In parallel with MCP's development, 11 major companies including Google, Microsoft and GitHub published the Agentic Resource Discovery (ARD) specification. ARD tackles an adjacent problem: how AI agents can find available tools and services on the internet.
Why it matters
MCP going stateless is fundamental for the scalability and reliability of systems built on the protocol. Stateless servers can handle any incoming request independently of prior interactions, which greatly simplifies autoscaling and load balancing and removes the need for sticky sessions. For infrastructure teams that means more flexible, more resilient systems. For existing MCP servers, though, migration is mandatory, and for many companies that will be a serious piece of work.
The reworked, agent-oriented authorization and the deprecation of three core features are the main incompatibility points for older servers. Maintainers have explicitly framed the ten-week gap between RC and final release as migration time. The ARD spec, published jointly by Google, Microsoft and GitHub, solves tool discovery. Organizations will be able to publish an ai-catalog.json on their domains listing MCP servers, tools and agents. Registries will index those catalogs, letting agents find the resources they need through natural-language queries. GitHub has already wired an agent finder into Copilot, and Hugging Face into its Discover Tool.
What it means for vibe coding
For developers who practice vibe coding, the MCP changes and the arrival of ARD mean an evolution in how you interact with tools and services. A stateless MCP can simplify plugging AI agents into distributed systems, making them more predictable and easier to manage. Agents that can discover tools through ARD open the door to more autonomous, adaptive assistants that pick the resources they need for a task on their own.
Developers running MCP servers will need to migrate soon to stay compatible with version 2026-07-28. That may mean rethinking the authorization architecture and dropping some deprecated features. At the same time, ARD makes finding and integrating new tools easier for AI agents, which could produce richer, more capable vibe coding scenarios. An agent might not only write the code but also track down the libraries, APIs or even other AI services it needs to run it.