What is the MCP 2026-07-28 update?

It is the newest revision of the Model Context Protocol, published on 28 July 2026 — the biggest change since the standard launched. Its headline change makes MCP stateless, so servers can run on serverless and edge infrastructure, and it adds a governed extensions framework and stronger authorization.

What does a stateless MCP mean for my business?

It means each request is self-contained instead of relying on an open session. In practice, your MCP server can be deployed on cheap serverless functions, scaled without sticky connections, and placed behind standard caches and load balancers — lowering hosting cost and simplifying operations.

Do I need to rebuild my existing MCP integration?

If you run an older MCP server, yes — this is a breaking revision (the handshake, session header, and some features are removed or deprecated). But there is a minimum twelve-month deprecation window, so you can plan a migration rather than rushing it.

Is the new spec supported by Claude and other assistants?

Yes. Anthropic has announced support for the 2026-07-28 specification in Claude, and as an open standard maintained publicly, other MCP-compatible tools are expected to adopt it as well.

The New MCP 2026-07-28 Update: A Stateless, Cloud-Ready Standard for Connecting AI to Your Systems

On 28 July 2026, the team behind the Model Context Protocol (MCP) published the 2026-07-28 revision of the specification — the largest change since the standard launched. In one line: MCP is now stateless. The protocol drops its old connect-and-keep-a-session design in favor of a simple request/response model, which means an MCP server can now run on ordinary serverless and edge infrastructure like any other web API. For any business connecting AI assistants to its own systems and data, this is the update that makes those integrations cheaper to run and far easier to scale.

A quick refresher: what MCP actually is

MCP is the open standard that lets an AI model talk to your tools and data through a single, consistent interface — your CRM, your inventory database, your booking system, your files. Instead of building a custom bridge for every model and every tool, you expose an MCP server once, and any MCP-compatible assistant (including Claude) can use it. We covered the basics in our MCP explainer and its rise as an open standard. The 2026-07-28 revision does not change that promise — it re-engineers the plumbing underneath it for production.

The headline change: MCP is now stateless

Previously, a client and server had to perform an initialize handshake and hold an open session, tracked with a special session-ID header. That worked on a laptop but was painful in the cloud: it forced sticky connections and made serverless deployment awkward. The new spec removes the handshake and the session header entirely. Every request now carries its own protocol version, client identity, and capabilities inside a small metadata block, so each call is fully self-contained.

The practical payoff is real: MCP servers can be deployed on serverless functions and edge networks, scaled horizontally without session affinity, and placed behind standard load balancers and caches. The revision also adds a server/discover call for clean version negotiation, a single long-lived subscriptions/listen stream for change notifications, and per-request caching hints (a freshness time and a public/private scope) that cut needless polling and improve response caching.

Extensions become a real framework — Apps and Tasks

The old spec bundled everything into the core. The new one keeps the core small and moves optional capabilities into a governed extensions framework, each with a reverse-DNS identifier and its own version, negotiated through a capabilities map. Two extensions matter most for business use:

MCP Apps: servers can now ship interactive HTML interfaces that the host renders inside a sandboxed iframe. An AI assistant can present a real form, a table, or a small dashboard — not just text — while keeping it security-reviewed and cached ahead of time.
Tasks: promoted from an experimental feature to an official extension after real production use. Tasks give agents a clean way to run long, multi-step work in the background — the client checks progress with tasks/get, sends input with tasks/update, and can cancel — instead of holding one blocking call open.

Stronger, enterprise-grade authorization

Security got a meaningful upgrade. Clients must now validate the iss (issuer) parameter on authorization responses per RFC 9207, credentials are firmly bound to the authorization server that issued them, and the fragile Dynamic Client Registration flow is deprecated in favor of Client ID Metadata Documents. For a business connecting AI to sensitive systems, these are the details that make an integration safe to put in front of an auditor.

What breaks — and the 12-month migration window

This is a breaking revision, so honesty matters: existing MCP servers will need work. The Roots, Sampling, and Logging features are deprecated, the old HTTP+SSE transport is now deprecated in favor of Streamable HTTP, and stream resumability is gone (a dropped stream means re-issuing the request). The good news is the standard now has a formal feature lifecycle with a minimum twelve-month deprecation window, so nothing disappears overnight. Anthropic has already announced support for the 2026-07-28 spec in Claude, which signals fast ecosystem adoption.

What this means for your business

If you are already building AI features, the takeaway is positive: MCP is now a cloud-native, production-ready foundation. Serverless deployment lowers hosting cost and scales on demand; versioned extensions let you add rich interfaces and long-running agents without waiting for the core spec; and the tighter auth model fits enterprise and regulatory expectations. If you have an older MCP integration, plan a migration within the window rather than rushing it. And if you have not connected AI to your systems yet, this is a good moment to start on the new foundation instead of building on the old one.

How Origami helps

As a technology company, we design and build MCP integrations that let AI assistants work securely with your real systems — your database, your ERP, your booking and support tools. We build to the new 2026-07-28 spec: stateless servers you can host on serverless or edge infrastructure, extensions for interactive interfaces and background tasks where they add value, and authorization set up correctly from day one. If you already run an MCP server on the old design, we can plan and execute the migration without disrupting your operations.

Conclusion

The 2026-07-28 revision is MCP growing up — from a clever local protocol into a distributed, cloud-ready standard for connecting AI to business systems. Stateless servers, a real extensions framework, and stronger authorization make it easier and safer to build AI into the tools you already run. Whether you are starting fresh or upgrading, the direction is clear, and building on the new foundation now will save you rework later.

The New MCP 2026-07-28 Update: A Stateless, Cloud-Ready Standard for Connecting AI to Your Systems