A new specification released July 28, 2026, aims to simplify how artificial intelligence systems connect to external tools and data sources. Dubbed MCP 2026-07-28, the standard introduces a stateless architecture, formalizes extension mechanisms, and tightens authentication requirements — three changes that developers say could reduce integration headaches and security risks.
What the stateless architecture changes
Under the new model, each AI-to-tool interaction is treated as an independent request. The AI doesn't hold onto session data between calls, which means less state management for developers and fewer opportunities for data to leak across requests. The spec defines how context is passed in each message rather than relying on a persistent connection or stored session.
That shift is meant to make integrations more predictable. If a request fails, there's no half-open session to clean up. The AI simply retries with the same context. The stateless design also makes it easier to scale — any server can handle any request without needing to know what happened before.
Standardized extensions for broader compatibility
MCP 2026-07-28 also codifies a way for third-party tools to extend the protocol without breaking existing implementations. The standard defines a registry for extension identifiers and a format for declaring capabilities. That means a plugin for a database, a file system, or an API gateway can advertise what it does in a machine-readable way, and the AI can discover and use it without custom code.
The extension system is optional — a basic MCP client can ignore extensions it doesn't understand — but the hope is that it will encourage a marketplace of compatible tools. The specification includes examples for common use cases like reading files, querying databases, and calling web APIs.
Hardened authentication and authorization
Security gets a significant upgrade in this release. The new standard requires all MCP connections to use mutual TLS or OAuth 2.0 with proof-of-possession tokens. That means both the AI client and the tool server must prove their identity before any data is exchanged.
The spec introduces scoped permissions. A tool can declare exactly what actions it allows — read-only, write, admin — and the AI must request those specific scopes. The authentication handshake now includes a capability negotiation step where both sides agree on what's allowed before any request is processed.
The changes address a long-standing concern in the AI integration space: how to prevent a compromised tool from being used to exfiltrate data or execute unauthorized commands. By hardening the authentication layer, the standard aims to make it harder for attackers to impersonate either side of the connection.
What comes next
The MCP working group has opened a 60-day public comment period ending October 1, 2026. After that, the specification will be finalized and reference implementations will be published. Several major AI platforms have already announced plans to support MCP 2026-07-28 in their next SDK releases, though none have committed to a specific date.
For developers currently building AI integrations, the transition will require updating client libraries and server adapters to handle the new authentication flow and stateless message format. The working group has published migration guides for the most popular programming languages.
