MCP Just Had Its Biggest Rewrite Since Launch. Your AI Tools Run on It.

July 30, 2026. On July 28 the Model Context Protocol shipped version 2026-07-28, the fifth and by far the largest revision of the standard since it launched. Most business owners have never heard of MCP. Almost all of them are already running on it. It is the wiring that lets an AI assistant reach the tools you pay for, and the list of things now speaking it includes GoHighLevel, HubSpot, Clay, Figma, Supabase, Zoom, Intuit and Xero. Anthropic, which co-created the protocol and published its own adoption note the same day, says Claude alone now lists more than 950 MCP servers in its connector directory.

This release matters because it breaks things on purpose, and then promises not to do that again for at least a year.

What shipped

1. The protocol went stateless. The initialize and initialized handshake is retired, along with the Mcp-Session-Id header. Every request now carries its own protocol version, client identity and capabilities, so any request can land on any server instance behind an ordinary round robin load balancer. There is a new optional server/discover call for clients that want a capability list up front.
2. Routing moved into HTTP headers. Streamable HTTP requests must now include Mcp-Method and Mcp-Name. Your gateway, rate limiter or firewall can route, authorize and meter on those headers instead of parsing the request body.
3. Mid task confirmations got a standard shape. Multi Round-Trip Requests replace the old server initiated elicitation, sampling and roots calls that needed a connection held open. The server answers with a result type of input required plus the questions it needs answered, and the client retries the original call with the answers attached.
4. Tool lists became cacheable. Responses from tools, prompts and resources listings now carry a time to live and a cache scope, so clients can stop re fetching the same catalog and can keep upstream prompt caches stable across reconnects.
5. Extensions became formal. Tasks joins MCP Apps and Enterprise-Managed Authorization under a versioned extensions framework. Tasks moved out of the experimental core into its own extension with a poll based get and a new update call, so long running work has an official pattern instead of a per vendor hack.
6. Authorization was hardened. Authorization servers should return the issuer parameter per RFC 9207 and clients must validate it before redeeming a code, client credentials are now bound to the issuer that minted them, and Dynamic Client Registration is formally deprecated in favour of Client ID Metadata Documents. Anthropic frames the practical result plainly: MCP servers now line up with production OAuth 2.0 and OpenID Connect deployments, so they connect to identity systems like Entra or Okta without workarounds.
7. There is finally a deprecation policy. Twelve months minimum before anything removed. Roots, Sampling and Logging are deprecated and keep working for at least a year. The legacy HTTP and server sent events transport is deprecated with a year long offramp.
8. The SDKs landed the same day. TypeScript, Python, Go and C# all speak the new version now, with Rust in beta.

The scale nobody talks about

Two numbers were published on July 28 by two different parties, and both are worth writing down. The MCP maintainers report close to half a billion downloads a month across the Tier 1 SDKs, with the TypeScript and Python SDKs each passing one billion total downloads. Anthropic, counting its own way, says MCP recently passed 400 million monthly SDK downloads, a four times increase this year. Take the smaller figure and the conclusion does not change. This is now infrastructure.

The most telling data point came from a supporting quote rather than the announcement. Austin Parker at Honeycomb said that nearly 20 percent of all monthly interactive queries on their platform are now made by agents rather than people. That is the shape of the change. Software your team bought for humans is increasingly being driven by machines, and the protocol under that traffic just got rebuilt for scale.

Who already supports it

Adoption on day zero was unusually broad. Amazon says the stateless core is available in Bedrock AgentCore, with the Tasks extension contributed by AWS itself, and AgentCore Gateway users can opt in by updating the gateway with the versions they want to support. Cloudflare shipped support in its Agents SDK and across its own MCP servers on release day, so servers can run in Workers with no transport session at all. Microsoft Foundry, Google Cloud, Netlify, FastMCP, PostHog and Supabase all published support or commitments alongside the release. Supabase named the change that mattered to them: their MCP server runs statelessly, so until now it could not ask a user to confirm a destructive query before running it. Now it can.

What it means for operators

You are not going to implement a protocol revision. Your vendors will, and the effects will arrive in your account without an announcement. Four of them are worth planning around.

Reliability and cost improve quietly. Stateless servers can sit behind ordinary load balancers and run on serverless infrastructure, which means fewer of the dropped connections that make an AI automation fail halfway through a run, and less of the infrastructure overhead that vendors eventually price into your bill.

Human approval becomes a built in step rather than a custom one. The confirm before you act pattern, the one that stops an agent refunding a customer or deleting a record without a person seeing it, is now part of the standard. If you have been avoiding automation because you could not put a human in the loop cleanly, that objection is weakening.

Long running work gets a shape. The Tasks extension means a job that takes twenty minutes can report progress and be cancelled through a documented interface instead of a polling script somebody wrote once and never documented. That is the difference between an automation you can hand to a client and one only its author can operate.

Access control moves to your identity provider. With authorization aligned to standard OAuth and OpenID Connect, the right question stops being which staff member connected which tool and becomes which group in your directory is allowed to reach which system. We covered that shift when the Enterprise-Managed Authorization extension went stable in June, and this release makes it the default direction of travel.

The twelve month clock is the real story

Agent infrastructure has never had a predictable upgrade calendar. Until now the pattern was that something changed, an integration broke, and somebody found out from a client. A formal twelve month deprecation window is the first time this layer behaves like the rest of the web stack, and it turns a category of surprise into a calendar entry.

Use it. Anything you own that depends on Roots, Sampling, Logging, Dynamic Client Registration or the legacy HTTP and server sent events transport has roughly a year, and you now know that in advance rather than after an outage.

What to do this week

1. List every AI connector running in the business. The ones inside your CRM count. So do the ones a contractor set up in a chat client.
2. For each one, ask the vendor which spec version their server speaks and when they will support 2026-07-28. A vendor who cannot answer that is telling you something.
3. Check whether any custom connector you commissioned uses Dynamic Client Registration. It still works, and it is now on a clock, so budget the change rather than discovering it.
4. Move connector access behind your identity provider so joiners and leavers are handled in one place instead of tool by tool. Agencies running client work should be doing this per client, not per tool.
5. Give every connector a named owner. If nobody owns it, nobody will notice when it stops working, and stateless or not, things still stop working.

None of that requires an engineer. Deciding what to migrate first probably does, which is where an AI engineer or an AI automation agency earns the fee. If your client automations run on GoHighLevel, the same audit applies to its MCP server, and if they run on n8n, your n8n developer should be pinning versions rather than tracking latest.

The honest caveats

Support is rolling out, not finished. Anthropic says support is coming to Claude products soon rather than today. The maintainers are explicit that the breaking changes carry a migration cost, particularly for anyone who built on session identifiers. And most of the ecosystem quotes in the announcement come from companies that helped test the release, which is normal for a launch and worth remembering when reading them. What is not in doubt is the direction. The connection layer under your AI tools is being rebuilt as ordinary, boring, scalable web infrastructure, and boring is exactly what you want underneath a system that talks to your customers.