MCP's Biggest Spec Update Yet: What Changes for AI Agents in 2026

The Model Context Protocol (MCP) — the connector standard that lets Claude, ChatGPT, Cursor, Zapier, and dozens of other AI tools plug into your apps and data — just shipped its largest specification update since it launched. Published July 28, 2026, the 2026-07-28 spec rips out MCP's stateful session model and replaces it with a plain request/response design, alongside hardened authorization and a formal system for adding new capabilities without breaking existing servers.

If you don't build MCP servers yourself, this can look like inside-baseball protocol news. It isn't. MCP has quietly become the plumbing behind most "connect your tools to AI" features shipping across the industry, so this update changes how reliably those integrations run, how cheaply vendors can host them, and how fast new AI-agent connectors get built. Here's what actually changed and who it affects.

Quick Take

• What it is: The 2026-07-28 release of the Model Context Protocol spec, maintained by the Agentic AI Foundation (a Linux Foundation project Anthropic donated MCP to).
• The headline change: MCP is now stateless at the protocol layer — servers no longer have to hold open sessions, so they can run behind ordinary load balancers instead of needing sticky routing and shared session stores.
• Also new: Hardened OAuth 2.1/OIDC-based authorization, a formal extensions framework for adding features without breaking the core spec, and a 12-month minimum deprecation policy before anything is removed.
• The catch: Existing MCP servers built on the old bidirectional, session-based transport aren't automatically compatible — maintainers have a migration window, but "always-on stateful connection" is no longer how the protocol is designed to work.

Why a Protocol Update Is Actually News for SaaS Buyers

MCP is the reason a growing list of AI products can say "connect your Slack, GitHub, or Notion" without each vendor writing custom integration code for every tool. Anthropic introduced it in late 2024 and later donated it to the Agentic AI Foundation; since then it's been adopted broadly enough that OpenAI, Google, Microsoft, and Amazon all ship MCP support in their own agent platforms, alongside automation tools like Zapier and n8n that build entire product lines on top of it.

That popularity was also becoming MCP's problem. The original protocol kept a live, bidirectional connection open between client and server to support features like server-initiated prompts. That's fine for a local tool running on your laptop; it's a genuine operational headache for a company trying to run thousands of MCP server instances in production, since it forces sticky load-balancer routing and a shared session store just to keep track of who's talking to whom.

What the 2026-07-28 Spec Actually Changes

According to the Model Context Protocol's official specification announcement, four changes matter most:

A stateless core. Every request can now be handled independently, so MCP servers scale on standard round-robin load balancers and ordinary Kubernetes setups — no sticky sessions, no dedicated Redis layer just to remember connection state. For the companies hosting these servers at scale, that's a real infrastructure cost saved, not just a technical nicety.

Multi Round-Trip Requests (MRTR). Features that used to require the server to reach back to the client mid-conversation — like asking the user a follow-up question — are redesigned around a request/response pattern: the server returns a structured "input required" result, the client collects the answer, and the original call is reissued with it. Same capability, no held-open connection required.

Authorization hardening. Remote MCP servers now have to use OAuth 2.1 with PKCE and RFC 8707 resource indicators (which stop a stolen token for one server from being replayed against another), plus issuer validation from RFC 9207. In practice, that means MCP servers plug into enterprise identity systems like Okta or Microsoft Entra without workarounds — a real gap for any business that was hesitant to expose internal tools through MCP before.

A formal extensions framework and deprecation policy. New capabilities can now ship as versioned extensions with their own maintainers, instead of every feature having to go through the core spec. And nothing gets removed from MCP with less than 12 months' notice — a first for a protocol that's changed fast enough to catch developers off guard before. Three older features (Roots, Sampling, and the original Logging design) move into deprecated status under this policy; they keep working during the transition window rather than breaking immediately.

Who's Already Moving

Anthropic is rolling the new spec directly into Claude, and used the moment to also introduce MCP tunnels (a research preview) that let Claude reach MCP servers sitting behind a company firewall without exposing them to the public internet — useful for enterprises that want Claude or Claude Code talking to internal tools without opening inbound network rules. Coverage from The Register and VentureBeat both frame this as the moment MCP stops being a convenient local-tool protocol and starts being treated as production infrastructure, which tracks with Microsoft, Google, and AWS all publishing their own migration notes for the new spec around the same release window.

For builders working in Cursor, GitHub Copilot, or Claude Code, the practical effect is mostly invisible short-term: your editor's AI agent already talks to MCP servers, and client-side SDK updates should handle most of the transition. The people who need to act now are the ones building and hosting MCP servers — anyone running an integration for a SaaS product needs to check whether they're on the old stateful transport, since new implementations shouldn't adopt it going forward even though it isn't pulled immediately.

What This Means If You're Choosing Tools, Not Building Them

If you're evaluating AI tools for your business rather than writing MCP servers, the direct impact is smaller but still real. Vendors that host a lot of MCP-based integrations — automation platforms, AI coding tools, and enterprise AI assistants — should get more reliable connector uptime and faster rollout of new integrations once they're on the new spec, since a stateless architecture is simply easier to run at scale without the sticky-session failure modes that plagued earlier deployments. The authorization hardening is the part worth asking vendors about directly if you're deploying AI agents inside a regulated or security-conscious organization: a vendor that's already moved to OAuth 2.1 with resource indicators is meaningfully harder to compromise via token replay than one still running the older auth model.

Frequently Asked Questions

What is MCP and why should I care about a spec update? MCP (Model Context Protocol) is the open standard that lets AI assistants like Claude and ChatGPT, plus coding tools and automation platforms, connect to your apps and data through a common interface instead of custom one-off integrations. Its July 28, 2026 update changes how reliably and securely those connections run — even if you never touch the protocol directly.

Do I need to do anything if I just use AI tools, not build them? No immediate action is required. Client applications like Claude, Claude Code, and Cursor handle the protocol details; the migration work falls on whoever hosts the MCP servers behind the integrations you use.

What actually breaks with this update? Servers built on the old always-open, session-based transport aren't automatically compatible with the stateless model. The spec's new 12-month deprecation policy means nothing is removed abruptly, but new servers shouldn't be built on the deprecated Roots, Sampling, or original Logging design going forward.

Who maintains MCP now? The Agentic AI Foundation, a Linux Foundation project, after Anthropic donated the protocol following its original 2024 launch. OpenAI, Google, Microsoft, and Amazon have all since shipped their own MCP support.

Is this related to Claude Opus 5? No — this is a protocol-level change that applies across the whole MCP ecosystem, not a model release. Anthropic is adopting the new spec in Claude around the same time it shipped Claude Opus 5, but the two are separate pieces of news that happen to have landed in the same week.

Bottom Line

MCP's 2026-07-28 update is a plumbing change, but it's the kind that determines whether the "connect any tool to AI" promise actually holds up at scale. Making the protocol stateless removes a real operational headache for anyone hosting MCP servers in production, the authorization overhaul closes a legitimate enterprise security gap, and the new deprecation policy means future changes should be less disruptive than this one. Most everyday users of Claude, Cursor, or Zapier-style automations won't notice a thing changed — which, for a piece of infrastructure this widely adopted, is exactly the point. For a broader look at where the AI-agent tooling market stands right now, see our roundup of developer tools worth using in 2026.