Claude’s MCP Goes Stateless: What Changed in the Major July 28, 2026 Update?

July 29, 2026Updated: July 29, 2026Articles

Claude’s developer account has announced a major update to MCP, the system that connects AI to external services. The specification is called “MCP 2026-07-28.”

The headline change is that MCP servers no longer need to remember a particular connection in order to handle requests. That may sound technical, but for everyday users it means MCP-powered services should become easier to scale and deploy in more places.

This article starts with what MCP actually does, then explains what the update changes, whether Claude users need to adjust anything right now, and what MCP server developers should check.

Here are the posts announcing the MCP update

The announcement came from Claude’s official developer account.

MCP is a common connector that lets AI use external tools

When you ask Claude to find a document in Google Drive or check an issue on GitHub, Claude needs a way to connect to those services.

MCP is a set of shared rules for making that connection. An MCP server acts like an interpreter between the AI and an external service, defining which data can be read and which actions can be taken.

It is similar to USB-C. Devices from different manufacturers can use the same connector when they follow the USB-C standard. In the same way, an MCP server can be used from Claude, ChatGPT, and other AI tools that support MCP.

So this update is not the launch of one new AI service. It is an update to the foundation that lets AI use outside tools.

The biggest change: servers no longer need to remember the connection

Here is a simple analogy: package delivery.

Older remote MCP setups were like sending every package from one customer to the same local depot. If a later package went to another depot, that depot would not know what happened before.

That meant operators needed sticky sessions to keep a user routed to the same server, plus a shared system for passing session information between servers. As the number of users grew, the operational burden grew too.

With MCP 2026-07-28, each request carries the information needed to process it. Any server instance can handle the request. This is what “stateless” means here: the protocol does not rely on hidden connection state.

Operators can now distribute requests across available servers more easily. Adding servers, deploying in multiple regions, or using environments such as Cloudflare Workers and Vercel becomes more practical.

Stateless does not mean that an application cannot remember anything. If a shopping cart or long-running job needs state, the server can return an explicit identifier such as a cart ID or task ID and receive it again on the next request. The state becomes visible application data instead of hidden transport state.

Claude users may not notice an immediate change

You do not need to rush to change your Claude settings because of this announcement. The biggest benefits go to people who build or operate MCP servers, especially remote servers.

If you use a local MCP server through standard input and output (stdio) in Claude Code, your normal workflow does not suddenly change. The new design matters when you want to make that server available to many users or deploy it remotely.

For remote MCP connectors and servers hosted on the internet, provider updates should improve reliability and make scaling easier. Anthropic’s official article says support is rolling out across Claude products. Do not assume that every Claude surface supports every part of the new specification immediately.

Interactive tools, long-running jobs, and company logins become easier to support

The update is not only about where servers run. MCP now has a clearer framework for extensions, which broadens the ways AI and external services can work together.

MCP Apps: show an interactive interface inside the conversation

Until now, external services usually returned text or structured data. MCP Apps can provide screens such as search results, dashboards, and forms that a compatible Claude client renders inside the conversation.

The server provides the interface, and the AI client displays it inside a safer sandboxed frame. This could reduce the need to leave a chat just to filter results or confirm an action.

Tasks: hand off work that does not finish immediately

Some work takes time: converting many files, running a long investigation, or processing a large dataset. Tasks separate starting the work from receiving the result.

That makes it easier to check progress with a task ID and collect the result later, without keeping one connection open for the entire operation.

Enterprise-Managed Auth: manage company permissions centrally

Companies often want to connect MCP to their identity provider and single sign-on system instead of asking every employee to authenticate separately. Enterprise-Managed Auth provides a foundation for managing access to MCP servers centrally.

Stronger authentication does not make an unknown MCP server safe by itself. You still need to verify who operates the server and which permissions it requests.

For individual users, checking support is enough for now

These simple rules are a good starting point:

Only using local MCP: no need to change settings in a hurry
Using an official Claude connector: wait for the connector to adopt the new support
Hosting a remote MCP server: check SDK and specification compatibility
Managing MCP for a company: review authentication, permissions, and auditing

There is no reason to force an older server to work with a new specification using only client-side settings. Check whether the provider has released a compatible version and follow its migration instructions.

Connectivity and safety are separate questions

The more useful MCP becomes, the more it can do through external services. Some servers can create tasks, edit files, or delete data—not just read information.

Claude’s official help recommends checking the following when adding a custom connector:

Is the operator trustworthy?
Are the requested OAuth permissions necessary and limited?
Which tools are enabled for the conversation?
Can the server only read, or can it also write and delete?
Has the server’s behavior changed after an update?

An easier connection does not mean that permission review is automatic. Treat an MCP connection like handing an external service a key to part of your data and workflow.

For my workflow, watching the rollout is enough

I use Claude and Codex for article planning, research, specifications, and implementation support. I am not currently running a local MCP server continuously through Claude Code, so I do not need to rewrite my settings because of this announcement.

If I later connect Obsidian, GitHub, or WordPress to Claude through a remote MCP server, the stateless design will matter. After authentication and permissions are designed properly, it should be easier to move the server from a Mac to cloud or edge infrastructure.

That does not mean I would publish a custom MCP server to the internet immediately. I would first use a trusted connector to check exactly what Claude can read and which actions it can take. For a custom server, I would start with read-only tools before adding tools that write, delete, or send data elsewhere.

Summary

The biggest change in MCP 2026-07-28 is that connections are no longer tied to one particular server instance. Operators can scale servers more easily, and users should eventually see more reliable remote MCP services.

The same update also organizes MCP Apps for interactive interfaces, Tasks for long-running work, and Enterprise-Managed Auth for centrally managed company access.

If you simply use Claude day to day, this is not a “change your settings today” announcement. Local MCP users can wait for their provider’s update. Remote MCP operators should check their SDK, authentication, and session handling. That distinction makes a highly technical specification change much easier to understand.