MCP 2026-07-28 Release Candidate: Stateless Protocol Unlocks Enterprise MCP Integrations

The Model Context Protocol (MCP) just took its biggest leap forward since launch. The 2026-07-28 Release Candidate—finalizing July 28, 2026—introduces a stateless protocol core, first-class extensions, hardened authorization, and a formal deprecation policy. For enterprises building on MCP, this isn't just an incremental update; it's the architectural shift that makes production-grade, horizontally scalable MCP deployments practical.

The Headline: MCP Goes Stateless

The defining change in 2026-07-28 is the removal of protocol-level sessions. In the previous 2025-11-25 spec, every MCP connection required an initialize handshake and an Mcp-Session-Id header that pinned clients to specific server instances. This meant:

Sticky sessions at the load balancer
Shared session stores across server replicas
Deep packet inspection for routing

With 2026-07-28, all of that disappears. The handshake is gone. The session ID is gone. Every request is self-contained, carrying protocol version, client info, and capabilities in a _meta field. A new Mcp-Method header lets load balancers route on the operation without inspecting the JSON-RPC body.

The practical impact is immediate: A remote MCP server that previously needed sticky sessions and a shared Redis cluster can now run behind a plain round-robin load balancer, route traffic on Mcp-Method, and let clients cache tools/list responses via the new ttlMs and cacheScope fields.

For enterprises running Kubernetes, serverless functions, or any horizontally scaled infrastructure, this eliminates an entire category of operational complexity.

First-Class Extensions: MCP Apps & Tasks

Extensions existed before but had no formal governance. 2026-07-28 changes that with SEP-2133: extensions now have reverse-DNS identifiers, negotiated capability maps, independent versioning, and a dedicated SEP track from experimental to official.

Two official extensions ship with this release:

MCP Apps — Server-Rendered UIs

MCP Apps (SEP-1865) let servers ship interactive HTML interfaces that hosts render in a sandboxed iframe. Tools declare their UI templates upfront so hosts can prefetch, cache, and security-review them before execution. Every UI-initiated action flows through the same JSON-RPC protocol, audit trail, and consent path as a direct tool call.

For enterprises, this means rich, server-controlled interfaces for complex workflows—approval dialogs, multi-step forms, data visualizations—without custom client development.

Tasks — Long-Running Operations

The Tasks extension reshapes async work around the stateless model. A server answers tools/call with a task handle; the client drives it with tasks/get, tasks/update, and tasks/cancel. Task creation is server-directed: the client advertises the extension and the server decides when a call should run as a task.

This replaces the experimental Tasks API from 2025-11-25 with a lifecycle that works natively in stateless deployments—critical for enterprise workflows like report generation, data pipeline triggers, or human-in-the-loop approvals that span minutes or hours.

Authorization Hardening: Enterprise-Grade OAuth/OIDC

Six SEPs align MCP's authorization with real-world OAuth 2.0 and OpenID Connect deployments:

RFC 9207 iss validation (SEP-2468): Clients must validate the issuer on authorization responses, mitigating mix-up attacks prevalent in MCP's single-client, many-server pattern.
Dynamic Client Registration with application_type (SEP-837): Desktop/CLI clients no longer get defaulted to "web" and rejected on localhost redirect URIs.
Issuer-bound credentials & re-registration (SEP-2352): Credentials bind to the authorization server's issuer; clients re-register automatically when resources migrate.
Refresh token documentation (SEP-2207), scope accumulation (SEP-2350), and .well-known discovery (SEP-2351) close remaining gaps.

For enterprises with existing identity providers (Okta, Entra ID, Keycloak, Auth0), this means MCP slots into your auth infrastructure without custom adapters or workarounds.

A Formal Deprecation Policy: Stability Guarantees

The new feature lifecycle policy (SEP-2577) gives every feature an Active → Deprecated → Removed lifecycle with at least 12 months between deprecation and earliest removal.

In this release, three core features are deprecated (annotation-only, still functional):

Roots → Tool parameters, resource URIs, or server configuration
Sampling → Direct LLM provider API integration
Logging → stderr for stdio; OpenTelemetry for structured observability

This predictability is essential for enterprise planning. You can build on 2026-07-28 knowing exactly how long deprecated features remain supported and that future revisions won't break core capabilities without a year's notice.

Why This Matters for Enterprise MCP Integrations

The Hacker News discussion on the release candidate validates the enterprise impact. Glama (an MCP server gateway/registry) reported that a significant portion of their issues stemmed from session persistence requirements. With ~30% of 62,000+ indexed open-source servers actively maintained, the stateless shift dramatically lowers the barrier to running MCP servers at scale.

Key enterprise benefits:

Simplified infrastructure: No session stores, no sticky routing, standard HTTP caching
Serverless-ready: MCP servers now run natively on Cloudflare Workers, AWS Lambda, Vercel, Cloud Run
Gateway-friendly: Mcp-Method/Mcp-Name headers let API gateways route, rate-limit, and observe without body parsing
Observability built-in: W3C Trace Context propagation in _meta means distributed traces correlate across SDKs, gateways, and downstream services
Extensibility without fragmentation: New capabilities ship as opt-in extensions, stabilizing independently before (if ever) entering the core spec

Neurux + MCP: Enterprise-Ready Today

Neurux has supported MCP since the mcp-identity-forwarding release, with Admin → MCP Servers management for both SSE and Streamable HTTP transports. Our implementation already leverages:

Identity forwarding: Automatic signed JWT tokens with every tool call (user ID, email, role, workspace) so MCP servers scope responses per user
Encrypted credential storage: Authorization headers and secrets encrypted at rest, masked in UI
Per-workspace enablement: Connect sensitive MCP servers once, expose only to authorized workspaces
Agent-only enforcement: MCP tools only work in Agent chat mode, preventing accidental exposure

With the 2026-07-28 spec, Neurux's MCP integration gains:

Horizontal scaling for MCP server connections without session affinity
MCP Apps support for rich, server-rendered tool UIs in the Neurux chat interface
Tasks extension for long-running enterprise workflows
Hardened OAuth/OIDC alignment with your existing identity infrastructure

Ready to Deploy Enterprise MCP?

The MCP 2026-07-28 Release Candidate delivers the foundation enterprises have been waiting for: a protocol that runs statelessly on commodity infrastructure, an extensions framework where capabilities evolve independently, and a lifecycle policy that guarantees stability.

Neurux makes it simple to connect, manage, and govern MCP servers across your organization—with identity forwarding, per-workspace access control, and audit logging built in.

Contact our team to learn how Neurux can help you deploy MCP at enterprise scale—securely, scalably, and with full governance.

The next generation of open, interoperable, enterprise-grade AI tool integration starts now.