The Protocol That Connects AI to Everything Gets Rewritten on July 28 — and Most of the Internet Isn't Ready

On Tuesday, July 28, the Model Context Protocol ships the largest breaking change in its history. Every AI tool server on earth is about to be sorted into two groups: the ones built for what's coming, and the ones that just became legacy. Here's exactly what changes, why it happened, and what you need to do about it — in plain English.

There's a protocol most people have never heard of that now sits underneath almost every AI product you use. It's called the Model Context Protocol (MCP) — the open standard that lets Claude, ChatGPT, Cursor, Copilot, and thousands of AI agents actually do things: read your calendar, update your CRM, send an invoice, deploy your code. Anthropic released it in November 2024. Within a year, OpenAI, Google, Microsoft, and AWS had all adopted it, the SDKs were being downloaded roughly 97 million times a month, and the whole thing was donated to the Linux Foundation to become neutral internet infrastructure.

If AI is the new electricity, MCP is the outlet standard.

And on July 28, 2026, that outlet gets redesigned.

The MCP 2026-07-28 specification is, in the maintainers' own words, the largest revision since the protocol launched — and it contains breaking changes. Not "read the release notes" changes. Architectural changes. The kind that decide which platforms scale into the agent economy and which ones spend the next year rebuilding their foundations.

I run an MCP orchestration platform (0nMCP — 800+ tools across 48+ services, all served from the edge), so I've spent the last two months living inside the release candidate. This is the breakdown I wish someone had written for me.

What is the MCP 2026-07-28 spec update?

Short answer: MCP is going stateless. The protocol is throwing away the "phone call" model — where a client dials in, does a handshake, and gets pinned to one server for the whole conversation — and replacing it with a "letter" model, where every single request carries everything needed to answer it, and any server anywhere can pick it up.

That one sentence explains almost every change in the release. Let's walk through them.

The handshake is gone

Since launch, every MCP conversation began with an initialize ritual: client and server exchange versions and capabilities, the server issues a session ID, and every subsequent request must carry that ID like a wristband at a concert. The new spec deletes the whole ceremony. The initialize handshake is removed, the Mcp-Session-Id header is removed, and the information that used to be exchanged once now travels inside every request. A new server/discover method lets clients ask "what can you do?" whenever they want.

Sessions are gone — and that's the entire point

The session was the original sin of MCP at scale. A session pins a client to one specific server instance, which means production deployments needed sticky routing, shared session stores, and load balancers that inspect message bodies to figure out where to send things. That works fine on a laptop. It falls apart when you're serving millions of agent requests across a global fleet.

In the new model, any request can land on any server, behind a plain round-robin load balancer, with nothing shared. Servers that need memory across steps — a shopping cart, a multi-step workflow — do what web APIs have always done: hand the AI a ticket (a basket_id, a run_id) and let the model pass it back on the next call. The state becomes visible to the model instead of hidden in transport plumbing, which turns out to be more powerful, not less: the model can reason about the handle, compose it across tools, and pass it between steps of an agent workflow.

Requests become routable, cacheable, and traceable

Three quieter changes matter enormously for anyone running MCP in production. Every Streamable HTTP request must now carry Mcp-Method and Mcp-Name headers, so gateways and rate-limiters can route on the operation without cracking open the JSON body — and servers must reject requests where the headers and body disagree, which closes a whole class of smuggling tricks. Tool lists and resources now ship with ttlMs and cacheScope fields modeled on HTTP caching, so clients finally know how long a response stays fresh and whether it's safe to share. And distributed tracing key names are locked into the spec, so a single request can be followed from the chat window through the client, the server, and everything the server calls downstream.

Translation: MCP traffic just started behaving like real web traffic. CDNs, edge networks, and observability stacks can finally treat it as a first-class citizen.

Extensions become the new frontier

Instead of stuffing every new capability into the core spec, MCP now has a formal Extensions framework — and the first two official extensions tell you exactly where this is heading.

MCP Apps (SEP-1865, co-authored by Anthropic and OpenAI) lets a server ship actual interactive interfaces — buttons, forms, dashboards — that render inside the chat window in a sandboxed frame. Claude, ChatGPT, and VS Code already support it. The era of every AI answer being a wall of text is ending; tools are becoming apps that live in the conversation.

Tasks graduates from an experimental core feature into an extension redesigned for the stateless world: a server can answer a tool call with a task handle, and the client polls it. Long-running agent work — the report that takes ten minutes, the deployment that takes an hour — finally has a proper home.

Authorization gets hardened, and three old features get retired

Six separate proposals tighten how MCP handles OAuth and OpenID Connect — closing mix-up attacks, fixing the localhost-redirect problem that plagued desktop clients, and formalizing refresh tokens. Meanwhile Roots, Sampling, and Logging are formally deprecated under a new lifecycle policy that guarantees at least twelve months before anything deprecated can be removed. Which brings us to the most underrated line in the entire release: MCP now has a rulebook for how it changes. Deprecation windows. A conformance suite that proposals must pass before going final. An extensions track where ideas prove themselves before touching core. This release breaks things precisely so future releases won't have to.

Why is MCP going stateless?

Because MCP won.

The protocol was designed in 2024 for a developer wiring tools to a chatbot on one machine. Then it became the connective tissue of the entire AI industry — running in production at companies of every size, powering agent fleets, mediating billions of tool calls. The stateful design that was elegant on a laptop became the single biggest obstacle to running MCP like internet infrastructure.

The maintainers had a choice: keep patching around sessions forever, or take one clean, painful break and land on an architecture that scales like the web itself. They chose the break. It's the right call — and it's also a sorting event.

Who wins and who loses on July 28?

Here's the part nobody says out loud.

If you built your MCP server the "session way" — stateful, sticky, holding conversation memory in server RAM — you now have a genuine migration project: rearchitecting state, rebuilding transport handling, retesting everything, all while the deprecation clock runs.

If you built stateless from day one — serverless, edge-deployed, every request self-contained — the new spec isn't a migration. It's the industry converging on your architecture. Cloudflare Workers, Lambda-style functions, edge runtimes: these were already letter-model systems living in a phone-call world. On July 28, the phone-call world ends.

I'll be transparent about my bias here: 0nMCP runs entirely on Cloudflare's edge, stateless since the first commit, which means our July 28 work was a tune-up measured in days while much of the ecosystem faces surgery measured in quarters. But the bias doesn't change the physics. Statelessness is how the web scaled, it's how CDNs work, and now it's how AI tool infrastructure works. The platforms that internalized that early just got handed a head start.

The money layer: why this spec unlocks an economy

Now connect the dots to what happened four weeks before this release.

On July 1, 2026, Cloudflare opened the waitlist for its Monetization Gateway — a system that lets anyone charge for web pages, APIs, datasets, and MCP tools at the network edge, settled instantly in stablecoins over the open x402 payment protocol. An AI agent hits a paid tool, receives a machine-readable price, pays automatically from its wallet, and gets its answer — no signup, no credit card, no human. The x402 standard is governed by a Linux Foundation body whose launch members include Visa, Mastercard, Stripe, Google, Microsoft, AWS, and Shopify, and it has already processed on the order of 165 million transactions at a median price of about three cents.

Here's why the timing isn't a coincidence: you cannot meter, cache, and charge for traffic you can't route. The stateless spec — with its routable headers and cacheable responses — is what makes it possible for an edge network to stand in front of an MCP server and run a toll booth on individual tool calls. July 28 isn't just a protocol cleanup. It's the substrate for an agent economy where every tool call can be a transaction.

For thirty years, the internet monetized human attention. The next internet monetizes machine capability — priced per call, settled in seconds, negotiated between software. The spec landing Tuesday is the paving under that road.

MCP 2026-07-28 migration checklist

If you run an MCP server, here's the practical list, in order of pain-avoided-per-hour:

1. Audit for session dependence. Grep for Mcp-Session-Id and any per-connection state. This finding alone tells you whether you have a tune-up or a rebuild.
2. Handle the new headers. Accept and validate Mcp-Method and Mcp-Name on every request; reject header/body mismatches.
3. Read client info per-request. Version, client identity, and capabilities now arrive in _meta on each call. Implement server/discover.
4. Convert hidden state to explicit handles. Anything that spans calls becomes an ID the model passes back as a normal argument.
5. Add cache metadata. Put ttlMs and cacheScope on your list and read results.
6. Update error codes and schemas. The missing-resource code moves to the JSON-RPC standard -32602, and tool schemas can now use full JSON Schema 2020-12.
7. Stop building on Roots, Sampling, and Logging. They keep working for a year-plus, but they're on the way out.
8. Run dual-mode. Keep serving legacy clients while you enable the new era. The deprecation policy gives you room — use it, and never hard-cut.
9. Publish your server card. Put discovery metadata at /.well-known/mcp.json so registries and agents can find you without connecting. Reachability is about to be a ranking factor.

FAQ

Does my MCP server break on July 28? No. The spec goes final that day, but deprecated features remain functional for at least twelve months, and clients will adopt the new version gradually. What starts on July 28 is the clock — and the divergence between servers that speak the new era and servers that don't.

Do I have to rewrite my client? If you maintain an MCP client, yes, there's real work: per-request metadata, new headers, the removal of the handshake, and the retry-based flow for server prompts. Tier-1 SDKs shipped betas during the release-candidate window, so most of the lift is upgrading and adapting.

Is stateless worse for complex agent workflows? The opposite, in practice. Explicit handles make workflow state visible to the model instead of buried in transport metadata — the model can reason about a run_id, branch on it, and carry it across tools. Hidden session state was a convenience for servers and a black box for agents.

What are MCP Apps? An official extension letting servers render interactive UI — forms, pickers, dashboards — inside AI chat clients, sandboxed and audited through the same consent path as any tool call. It's supported today in Claude, ChatGPT, and VS Code, and it's the clearest signal of where AI interfaces are going: from answers to applications.

What is x402, in one sentence? A payment standard that revives HTTP's dormant "402 Payment Required" status code so AI agents can pay per request in stablecoins — no accounts, no checkout — and it's now backed by a Linux Foundation coalition including Visa, Mastercard, Stripe, Google, and Cloudflare.

Where do I read the primary sources? The official announcement and changelog live at blog.modelcontextprotocol.io and modelcontextprotocol.io/specification. Cloudflare's Monetization Gateway announcement is at blog.cloudflare.com/monetization-gateway. Read the originals — this space moves too fast for secondhand summaries, including this one.

The bottom line

Every infrastructure standard has one moment where it stops being a developer toy and becomes plumbing — the moment it gets boring, stateless, cacheable, and billable. HTTP had it. Containers had it. On July 28, MCP has it.

Most people will never notice. They'll just find that their AI assistant got faster, that tools started rendering real interfaces instead of text, and that software quietly began paying software for services rendered. But if you build in this space, mark the date. The agent economy needed a foundation that scales like the web. On Tuesday, it gets one.

Mike Mento is the founder of RocketOpp and the 0nMCP ecosystem — a universal AI orchestration platform serving 800+ tools across 48+ services from the edge, stateless since day one. He writes about the infrastructure layer of the agent economy.