MCP Protocol Gets Its Biggest Update Yet on July 28 With Breaking Changes as GitHub Joins the Adoption Wave

MCP Goes Stateless on July 28: What Breaks, What Migrates

The Model Context Protocol ships its largest specification revision on July 28, 2026. The release candidate has been available since May 1. Now the final spec arrives, and the redesign is unmistakable: MCP is removing its stateful session model and rebuilding around a stateless core designed to run on ordinary HTTP infrastructure 1.

If you have an MCP server in production, here is what the migration looks like.

Goodbye, handshake

Until now, every MCP connection opened with a mandatory initialize/initialized handshake. That exchange negotiated protocol version, swapped capabilities, and pinned the client to a specific server instance through an Mcp-Session-Id header 1. The model worked when MCP was a single AI application talking to a local server process over stdio. It stopped working the moment organizations deployed MCP servers behind load balancers or across Kubernetes clusters. Sticky sessions, shared session stores, reconnection logic after dropped connections: all operational overhead that has nothing to do with what MCP exists to solve 2.

The new spec removes the handshake entirely. Every JSON-RPC request becomes self-contained, carrying protocol version, client identity, and capability flags inside a _meta object on the request itself 1. Any server instance in a pool can handle any request. Standard round-robin load balancing works without special configuration 2.

Two new mandatory headers, Mcp-Method and Mcp-Name, travel on every Streamable HTTP request so that load balancers and gateways can route by operation without parsing the JSON body 1. Servers reject requests where the headers and the body disagree.

Servers that still need to track state across calls can mint explicit identifiers (a basket_id, a workflow_id) from a tool result and have the model pass them back as ordinary arguments. The maintainers argue this pattern is often more powerful than hidden session state, because the model can reason about and compose identifiers in ways externally managed metadata never allowed 1.

Extensions get a real framework

The previous spec mentioned extensions but gave them no formal lifecycle. The new release adds one. Extensions are now identified by reverse-DNS IDs, negotiated through a capabilities map on both client and server, housed in their own repositories with delegated maintainers, and versioned independently of the core specification 1.

Two extensions ship in this release. MCP Apps lets servers declare interactive HTML interfaces that hosts render in a sandboxed iframe, so a server can deliver rich UI instead of handing the client raw data to interpret 2. Tasks gives servers a durable pattern for long-running operations: the server returns a taskId, and the client polls for status and submits input through dedicated RPC methods 2.

Three features move to deprecation

Roots, Sampling, and protocol-level Logging enter a formal deprecation process. Roots let clients advertise filesystem paths to servers; the migration path is to pass paths as tool parameters or server configuration. Sampling let servers request LLM completions from the client; servers should now call LLM APIs directly. Protocol-level Logging is replaced by stderr or OpenTelemetry 2. The MCP blog confirms all three are deprecated in this release 1.

A new deprecation policy, SEP-2596, guarantees a minimum twelve-month window between deprecation and removal, with a public registry of deprecated features so implementers can plan migrations rather than getting surprised 2.

Authorization also tightens. The release candidate requires OAuth 2.1 and OpenID Connect, replacing the basic OAuth support in earlier versions, and adds an Enterprise-Managed Authorization extension for organizations that need centralized identity management 2.

The graduation

MCP was built for a laptop. The persistent session, the stdio transport, the initialize handshake: every original design decision assumed one application talking to one local process 2. The July 28 spec is the protocol's response to what happens when that assumption meets load balancers, API gateways, multi-region deployments, and enterprise security requirements.

The breaking changes are real but bounded. Removing the handshake, adding the mandatory Mcp-Method and Mcp-Name headers, and replacing session-time capability negotiation with on-demand server/discover calls all require code changes in existing servers 1. The twelve-month deprecation window means Roots, Sampling, and Logging will not disappear overnight 2.

The bet the maintainers are making is straightforward. A cleaner, more scalable core earns its migration cost if the ecosystem migrates rather than forks. If enough server operators and tooling vendors adopt the new spec, MCP consolidates into something like the USB-C of AI agent connectivity: one interface shape every tool and every agent can rely on. If implementers balk at the breaking changes and fragment instead, the protocol loses the momentum it has spent the last year building. July 28 is when that question gets asked.