Protocols & news

The new MCP spec (2026-07-28): stateless servers and what changed

Listen · 10 min

Narrated by an AI voice. Code examples are on the page.

Short answer

The 2026-07-28 MCP spec makes the protocol stateless. It removes sessions and the initialize handshake, so every request carries its own version and client details, and it adds a required server/discover call, cache hints on tool lists and new routing headers. Clients built only for the old handshake cannot talk to a server that supports only the new version, so most servers will need to support both eras for a while.

Key takeaways

  • MCP servers no longer hold sessions; each request stands on its own.
  • Old handshake-only clients fail against new-only servers, so plan a dual-era period.
  • Roots, Sampling and Logging are deprecated, with at least twelve months' notice.

What changed in the MCP spec on 28 July 2026?

The 2026-07-28 MCP spec turned the Model Context Protocol into a stateless, request-and-response protocol. A server no longer keeps a session for each connected AI assistant. Every request now carries everything the server needs to answer it.

The MCP project's release post puts it this way: "MCP is transforming from a bidirectional stateful protocol into a request/response stateless protocol." It calls the stateless core the highlight of the release and says it was one of the most requested features from developers who wanted more reliable, scalable servers.

The release followed a release candidate published on 21 May 2026. The MCP project said its TypeScript, Python, Go and C# SDKs were updated to match.

If you are new to MCP itself, start with what MCP is. This article covers only what the 2026-07-28 revision changed and what that means if you run, or pay someone to run, an MCP server.

Why did MCP get rid of sessions?

Because sessions made remote servers hard to scale. Under the 2025-11-25 spec, a client opened a session with an initialize handshake, and later requests had to reach the server that held that session.

Google, which says it co-founded the MCP Transports Working Group, wrote in August 2026 that the old session model "required persistent state, handshakes, and session pinning." For a large hosted service, that means sticky load balancing or a shared store just to keep track of who is connected.

The new spec removes all of that:

  • No handshake. The initialize and notifications/initialized exchange is gone.
  • No session header. The Mcp-Session-Id header is removed from the Streamable HTTP transport.
  • Self-describing requests. Each request carries its protocol version and the client's capabilities in a _meta field, and clients should identify themselves on every request.

The MCP release post says the result is that "Any request can now land on any server instance behind a plain round-robin load balancer without needing shared storage."

Stateless does not mean forgetful. The post adds: "Dropping the protocol-level session doesn’t force your application to be stateless." A tool that needs to remember something, such as a cart, can return a handle that the AI model passes back on the next call.

What new parts does a server need to support?

Four changes matter most for anyone running a server.

1. server/discover. Servers must now implement this call. It returns the protocol versions, capabilities and identity of the server in one request. Clients may call it first, but they do not have to. A client can send any request directly and handle an UnsupportedProtocolVersionError if the version does not match.

2. Routing headers. Streamable HTTP requests must now include Mcp-Method and Mcp-Name headers, naming the method and, for example, the tool being called. The release post says "Your gateway, rate limiter, or WAF can route and meter on those headers instead of parsing JSON bodies." For a store, that makes it easier to rate-limit an expensive tool without inspecting every request body.

3. Cache hints on lists. Results from tools/list, prompts/list, resources/list, resources/read and resources/templates/list must carry a ttlMs freshness hint and a cacheScope of public or private. Servers should also return tools in a fixed order, which helps clients cache them.

4. Multi Round-Trip Requests. A server used to send its own requests to the client mid-call, for example to ask the user a question. Now it returns an input_required result listing what it needs, and the client retries the original request with the answers attached. The spec calls this "a breaking change." For a shop, this is how a tool can still ask a shopper to confirm a size or an address in the middle of an order.

Change notifications also moved: the old HTTP GET endpoint and resource subscriptions are replaced by a single subscriptions/listen stream that clients opt in to.

Will old AI clients still work with a new server?

Only if the server supports both versions. This is the most important practical point in the new spec.

The spec calls clients and servers that use the handshake "legacy" (2025-11-25 and earlier) and those that use per-request metadata "modern." Its compatibility matrix is blunt:

Client Server Result
Modern Modern Works
Legacy Modern only Fails
Modern Legacy only Fails
Legacy or modern Dual-era (supports both) Works
Dual-era Legacy or modern Works

The spec explains why a legacy client cannot recover on its own: "Legacy clients have no fall-forward mechanism." A dual-era server answers an initialize request the old way and serves modern requests statelessly, and it may do both at the same endpoint.

What this means for you:

  • If you run a store MCP server, do not switch it to modern-only until the AI clients your customers use support 2026-07-28. Supporting both is allowed and is the safe default.
  • If an agency or vendor runs it, ask which protocol versions the server accepts today, and whether it answers server/discover.
  • Test with real clients. The AI assistants that connect to your server may move to the new version at different speeds.

What was deprecated, and when will it be removed?

Three features are deprecated: Roots, Sampling and Logging. The changelog says "New implementations should not adopt them," but they "remain fully functional during the deprecation window."

The spec now has a formal deprecation policy. A deprecated feature stays in the spec for at least twelve months, or, with core maintainer approval, as little as ninety days if it poses an active security risk. The deprecated features registry lists the earliest removal for Roots, Sampling and Logging as the first revision released on or after 28 July 2027. The suggested replacements are tool parameters or resource URIs instead of Roots, calling AI model providers directly instead of Sampling, and standard error output or OpenTelemetry instead of Logging.

Other changes in this group:

  • Dynamic Client Registration is deprecated in favour of Client ID Metadata Documents, a way for an AI client to identify itself during sign-in. It still works for backward compatibility.
  • The old HTTP+SSE transport, deprecated since 2025, is now formally on the removal path. Streamable HTTP is the replacement.
  • Tasks, the long-running job feature added experimentally in the 2025-11-25 revision, moved out of the core protocol into an official extension.
  • Stream resumption is gone. If a response stream breaks, the client must send the request again with a new ID.

What changed for sign-in and security?

The revision tightens how AI clients sign in to protected MCP servers. This matters if your server can see orders or customer details.

  • Authorization servers should include the iss parameter defined in RFC 9207, and clients must check it before using a sign-in code. The release post says this "closes an authorization-server mix-up hole."
  • Client credentials are tied to the authorization server that issued them and must not be reused with a different one.
  • Clients must state their application type when they register, which fixes redirect errors for desktop and command-line apps.

None of this replaces your own controls. Keep tools that place orders or change data behind confirmation and limits, as our guide on whether your store needs an MCP server explains. For where MCP sits next to agent-to-agent and checkout standards, see MCP vs A2A vs UCP.

How Sill helps

Sill runs an MCP server for each verified site that lists your site's skills as tools, alongside llms.txt, an agent card and a catalog, and keeps those files current. That server supports the 2026-07-28 spec and the older handshake at the same address, so AI assistants on either version can connect while they move over. Every tool call is recorded in your audit log, and when an agent places an order, your rules approve it, decline it or hold it for a person to review. Run a free readiness scan to see whether agents can find an MCP server on your site today.

Questions people ask

Is MCP 2026-07-28 the current version?

Yes. At the time of writing, the MCP versioning page names 2026-07-28 as the current protocol version. The previous revision was 2025-11-25.

Do I have to rebuild my MCP server for the new spec?

Not straight away. The spec lets a server support both the old handshake-based versions and the new stateless one at the same endpoint. If you use an official SDK, upgrading the SDK is the usual route, and the MCP project says the migration has some cost for servers that relied on session IDs.

Does a stateless MCP server mean my tools cannot remember anything?

No. The protocol no longer keeps a session, but your application can. The MCP project suggests that a tool return an explicit handle, such as a cart ID, which the AI model passes back as an argument on later calls.

When will Roots, Sampling and Logging be removed?

Not before 2027. The deprecated features registry lists their earliest removal as the first spec revision released on or after 28 July 2027. They keep working until then.

Sources

  1. MCP specification 2026-07-28: Key changes · checked 17 Nov 2026
  2. MCP blog: The 2026-07-28 Specification · checked 17 Nov 2026
  3. MCP blog: The 2026-07-28 MCP Specification Release Candidate · checked 17 Nov 2026
  4. MCP specification 2026-07-28: Versioning and Compatibility · checked 17 Nov 2026
  5. MCP: Versioning · checked 17 Nov 2026
  6. MCP specification 2026-07-28: Deprecated features · checked 17 Nov 2026
  7. MCP: Feature lifecycle and deprecation · checked 17 Nov 2026
  8. MCP specification 2026-07-28: Multi Round-Trip Requests · checked 17 Nov 2026
  9. MCP specification 2025-11-25: Key changes · checked 17 Nov 2026
  10. Google Developers Blog: Scaling AI agent infrastructure with the MCP stateless updates · checked 17 Nov 2026

Written by The Sill Signals Team. Last reviewed 18 Nov 2026.

Share this articleLinkedInXBlueskyEmail

Keep reading

See which AI agents visit your site, and decide what they may do.

Scan your site free