The Model Context Protocol shipped its fifth specification release on July 28, and it reads like a protocol growing into its job. The 2026-07-28 spec rebuilds MCP around a stateless request/response core and an authorization model aligned with OAuth 2.0 and OpenID Connect practice. The release carries endorsements from executives at AWS, Google Cloud, Microsoft, and Cloudflare, and Anthropic rolled the changes into Claude the same day. The signal for operators: the connector layer between models and business systems is hardening into real infrastructure — the kind your load balancers, gateways, and identity provider already understand.
Key takeaways
- The core is now stateless. MCP moves from a stateful bidirectional protocol to plain request/response, designed to run on serverless and edge infrastructure behind standard load balancers.
- Authorization aligns with your identity stack. Six SEPs bring the auth model in line with OAuth 2.0 and OpenID Connect practice, including required issuer validation to mitigate mix-up attacks.
- Anthropic shipped it into Claude the same day. Enterprise-managed auth, observability dashboards for connector usage, and MCP tunnels for private-network access (research preview).
- The platform vendors are on board. Endorsements in the release post came from executives at AWS, Google Cloud, Microsoft, Cloudflare, and others; Tier 1 SDKs are updated for TypeScript, Python, Go, and C#, with Rust in beta.
- It is a breaking rewrite. Existing MCP servers need migration work. Budget for it now, not later.
What the stateless core changes
Until this release, MCP was a stateful, bidirectional protocol. Servers held sessions open, which meant sticky routing, shared session storage, and held-open streams — none of which fit standard enterprise HTTP infrastructure. The new spec removes the Mcp-Session-Id header from the Streamable HTTP transport entirely. An MCP server is now a plain request/response service that can sit behind a round-robin load balancer with no shared state.
Three mechanics replace what statefulness used to do:
- Multi round-trip requests. When a server needs more input, it returns a
resultTypeofinput_requiredand the client retries withinputResponses— replacing held-open streams. - Header-based routing. New
Mcp-MethodandMcp-NameHTTP headers let gateways route requests without parsing JSON bodies. - Cacheable lists. List results carry
ttlMsandcacheScopeparameters, so clients and gateways can cache tool listings instead of refetching them.
This is the shape of every other service in your stack, and that is the point. An MCP server you can deploy on serverless infrastructure, scale horizontally, and observe with tooling you already own is one your platform team can actually operate.
A protocol becomes infrastructure on the day your load balancer stops caring that it exists.
Authorization catches up to enterprise practice
Six SEPs harden the authorization model to match how OAuth 2.0 and OpenID Connect are actually deployed. Validation of the iss parameter is now required, closing off mix-up attacks. The spec adopts RFC 9207 issuer validation, deprecates Dynamic Client Registration in favor of Client ID Metadata Documents, and binds credentials to their issuer.
The release also formalizes an extensions framework. Tasks (long-running work), MCP Apps (server-rendered UI inside conversations), and Enterprise Managed Authorization are now official extensions rather than vendor experiments. That matters for procurement: an extension with a published spec is something you can require in a contract.
What Anthropic shipped in Claude
Anthropic rolled the spec into Claude alongside the release. The enterprise pieces: enterprise-managed auth, which supports org-wide connector provisioning through your identity provider; observability dashboards for connector usage; and MCP tunnels, a research preview designed to let agents reach tools on private networks without exposing public endpoints. That last one addresses a real objection — until now, giving an agent access to an internal system usually meant opening something to the internet.
On scale, the numbers are Anthropic's own: the company says Claude's connector directory lists 950+ MCP servers and that the protocol has reached 400 million monthly SDK downloads, up 4× year-over-year. Treat those as vendor figures. The endorsement list — AWS, Cloudflare, Figma, Google Cloud, Honeycomb, Microsoft, Netlify, Supabase, Xero — is the more telling signal, because those are companies that have to build against this spec.
What operators should do now
For teams running tailored systems, this release lowers two costs at once. Integration cost drops because MCP servers now fit the infrastructure you already run. Lock-in risk drops because the connector layer is a published spec with multi-vendor backing, not a per-platform SDK. Four moves for this quarter:
- 01Inventory your connector layer. List every MCP server you run or depend on, and which spec revision each one targets.
- 02Budget the migration. The stateless core and auth changes are breaking. Scope the work now rather than discovering it during an incident.
- 03Press your vendors. Ask every tool vendor with an MCP server for their timeline on the 2026-07-28 spec — and specifically on the new authorization model.
- 04Bring identity in early. Enterprise-managed auth moves connector provisioning into the identity provider. That is a security win, but only if your IdP team is at the table from the start.
In June we argued that agent interoperability standards would decide how much of an agent stack stays portable. This release is that argument landing. It also changes the build-vs-buy math on connector work, which sits at the center of how we scope custom AI applications. If you are planning an MCP migration — or a first deployment on the new spec — book a consult.