The Model Context Protocol (MCP) team has published a release candidate for the specification dated 2026-07-28, describing it as the largest revision of the protocol since its launch. The core is moving to a fully stateless architecture, a formal framework for extensions is arriving, and two headline additions – MCP Apps and Tasks – are changing how AI agents can talk to web applications and handle long-running work.
No more session IDs or handshakes
The new specification eliminates the initialize handshake and the Mcp-Session-Id header entirely – the mechanisms that used to keep client-server connections stateful. The result: any request can land on any server instance, which massively simplifies scaling MCP servers behind a load balancer. For companies currently running MCP integrations on a single server to avoid “sticky sessions,” this is a meaningful architectural relief – a stateless model enables horizontal scaling without special infrastructure tricks.
Extensions now have a life of their own
The spec introduces a formal extensions framework, identified by reverse-DNS IDs, negotiated through an extensions map, and living in their own repositories with delegated maintainers. That means individual extensions can version independently of the core specification and evolve at their own pace – not unlike how individual web standards are maintained under the W3C umbrella.
MCP Apps: interactive interfaces instead of plain text
The MCP Apps extension lets servers deliver interactive HTML interfaces that host applications render inside a sandboxed iframe. For developers building AI assistants and agents into client web projects today, that opens up an interesting option: instead of an AI agent returning only text or JSON, it can hand the user an interactive card, form, or data preview directly – without the host application having to hand-build custom UI for every single tool.
Tasks: no more waiting on long-running operations
The Tasks extension, which existed as an experimental core feature in earlier spec versions, has now “graduated” into a standalone extension with a stateless model for long-running work. A server can respond to a tools/call request with just a task handle, and the client then drives progress using tasks/get, tasks/update, and tasks/cancel. That’s crucial for the use cases we’re increasingly seeing in agency work – long-form content generation, bulk SEO audits, or WooCommerce automation that can take anywhere from minutes to hours and can’t afford to block on a single open connection.
What it means for your website right now
The final version of the specification is due to publish on July 28, 2026, with a ten-week window for SDK maintainers and client implementers to validate the changes against real workloads. If your agency or internal team is integrating AI agents into web projects – whether that’s chatbots, content-admin assistants, or e-commerce automation – it’s worth tracking how the MCP client and server ecosystem adapts to this revision, and designing new integrations with a stateless architecture in mind from day one.
A practical impact on SEO and content workflows
The stateless model behind the Tasks extension also has an interesting spillover into SEO and content marketing. Agencies routinely connect CMS platforms to AI tools for drafting articles, generating meta descriptions, or running bulk backlink audits – and it’s exactly these kinds of jobs, which can run for tens of minutes, that used to bump into timeout limits and force awkward request-queuing workarounds. A standardized way to kick off a task, monitor its progress, and cancel it if needed will make it considerably easier to build reliable automation on top of WordPress and other content platforms.
Security questions worth watching
The new ability to return interactive HTML interfaces directly from an MCP server naturally introduces new security considerations. Even though MCP Apps render inside a sandboxed iframe, teams building integrations should treat this content from day one as something generated and delivered by an external server – in other words, a potential attack vector if input validation and permissions aren’t handled rigorously. Before rolling integrations like this into a client’s production environment, it’s worth tracking how the major SDKs handle isolation between individual MCP servers and how they manage permissions for writing data back into the host application.
Conclusion
MCP’s move to a stateless core, paired with the ability to return interactive interfaces and safely handle long-running tasks, pushes the protocol closer to genuine production-grade infrastructure rather than an experimental technology. Thinking about bringing AI agents into your website or online store? We’re happy to help you make sense of what today’s tools can already do – and what’s still on the way.
