OpenToolslogo
ToolsExpertsNewsletter
AdvertiseLearn AI
  1. Home
  2. /
  3. News
  4. /
  5. MCP 2026-07-28: what the stateless revision breaks in your server or client, and when migrating is worth it
Updated 6 minutes ago

Share this article

PostShare

In This Article

  • What the stateless core changes
  • What the revision removes
  • Routing, caching and discovery additions
  • MRTR replaces the held-open-stream flows
  • Deprecations and the twelve-month window
  • Authorization hardening and the DCR-to-CIMD shift
  • SDK boundaries: TypeScript and Python verified, others qualified
  • Debugging the cutover with the Inspector's protocol eras
  • When migration is worth it

AI news in your inbox

Weekly updates on tools, models, and the companies building them.

Subscribe

Footer

Company name

The right AI tool is out there. We'll help you find it.

LinkedInX

Knowledge Hub

  • News
  • Resources
  • Newsletter
  • Blog
  • AI Tool Reviews
  • YouTube Summary
  • YouTube Transcript Generator

Industry Hub

  • AI Companies
  • AI Tools
  • AI Models
  • MCP Servers
  • Muse Connectors
  • AI Tool Categories
  • Top AI Use Cases

For Builders

  • Submit a Tool
  • Submit a MCP
  • Experts & Agencies
  • Advertise
  • Compare Tools
  • Favourites

Legal

  • Privacy Policy
  • Terms of Service

© 2026 OpenTools - All rights reserved.

MCP 2026-07-28: what the stateless revision breaks in your server or client, and when migrating is worth it
Source: Documentation screenshot from the Model Context Protocol project's 2026-07-28 protocol-eras page (Inspector Network tab, showing the 400 Bad Request the strict modern-HTTP server returned); repo commit 2be673cf. Credit: Model Context Protocol project. Licensed CC-BY-4.0. Image: Model Context Protocol project (per the licensing statement below); no individual author is named in the captured evidence.. CC-BY-4.0 (Creative Commons Attribution 4.0 International). · Source image · CC-BY-4.0 (Creative Commons Attribution 4.0 International)

MCP 2026-07-28: what the stateless revision breaks in your server or client, and when migrating is worth it

By OpenTools Team · November 16, 2026

The Model Context Protocol's 2026‑07‑28 revision turns MCP from a stateful, bidirectional protocol into a stateless request/response one. The initialize handshake and the Mcp‑Session‑Id header are gone; each request now carries its protocol version, client identity and client capabilities in _meta, and state that must survive across calls has to travel as an explicit, server‑minted handle passed as a tool argument. Server‑initiated roots/list, sampling/createMessage and elicitation/create are replaced by Multi Round‑Trip Requests: the server returns a result marked input_required with the requests it needs answered, and the client retries the original call with inputResponses. Streamable HTTP now requires Mcp‑Method and Mcp‑Name headers, list and read results carry ttlMs and cacheScope, and authorization is hardened around RFC 9207 issuer validation and a formal shift away from Dynamic Client Registration toward Client ID Metadata Documents. Roots, Sampling, Logging and the legacy HTTP+SSE transport are deprecated inside a minimum twelve‑month window. The TypeScript SDK splits into v2 packages while v1.x stays frozen at 2025‑11‑25, and the Python SDK is on a 2.3.0 line. The release post names Go and C# as day‑one Tier 1 SDKs and Rust as beta, but this evidence carries no version numbers for them, and the claimed performance gains are vendor statements rather than independent measurements.

What the stateless core changes

The 2026-07-28 revision moves MCP from a bidirectional, stateful protocol to a request/response core. The initialize and notifications/initialized handshake and the Mcp-Session-Id header are removed from Streamable HTTP, and each request now carries its protocol version, client identity and client capabilities in _meta; a version mismatch returns UnsupportedProtocolVersionError.1 Because list endpoints no longer vary per connection, a request can land on any server instance behind ordinary round-robin load balancing with no shared storage layer 1.2

Removing the transport session does not require your server to hold no state. The maintainers recommend minting an explicit handle from a tool and having the model pass it back as a normal argument, which they argue beats session state hidden in the transport because the model can see the handle and thread it between tools.2 The cost of ignoring that shows up in the Inspector's test setup: a stateless handler constructed per request cannot hold state between calls, so its modern subscription configuration omits a mutation tool that the legacy configuration keeps.3

What the revision removes

Treat this as the breakage checklist. The changelog removes, replaces or renames:

  • the initialize handshake and Mcp-Session-Id (SEP-2575, SEP-2567) 1
  • the HTTP GET notification endpoint, plus resources/subscribe and resources/unsubscribe, replaced by subscriptions/listen, a single long-lived POST-response stream that clients opt into per notification type 1
  • ping, logging/setLevel and notifications/roots/list_changed; log level becomes per-request through io.modelcontextprotocol/logLevel in _meta, and a server must not emit notifications/message for a request that did not opt in 1
  • SSE stream resumability and message redelivery, including Last-Event-ID and SSE event IDs: a broken response stream loses the in-flight request, and clients must re-issue it as a new request with a new request ID 1
  • the server-initiated roots/list, sampling/createMessage and elicitation/create requests 1
  • notifications/elicitation/complete and the elicitationId field added in 2025-11-25; servers that need to correlate an interaction across retries encode their own identifier in requestState 1

Error codes were renumbered too: resource-not-found moves from -32002 to -32602 (Invalid Params), HeaderMismatch from -32001 to -32020, MissingRequiredClientCapability from -32003 to -32021 and UnsupportedProtocolVersion from -32004 to -32022.1 One behavioural change is easy to miss: on modern Streamable HTTP, cancelling an in-flight tool call closes that request's own SSE response stream rather than sending notifications/cancelled, while stdio still uses the notification.3 The Inspector renders the -32602 family as distinct Unknown Tool and Invalid Parameters panels, reading the message text to tell them apart.3

Routing, caching and discovery additions

Streamable HTTP POST requests now must carry Mcp-Method and Mcp-Name headers, so gateways, rate limiters and WAFs can route and meter on headers instead of parsing JSON bodies (SEP-2243) 2.1 A tool can annotate an argument with x-mcp-header to ask a client to mirror that argument's value into an Mcp-Param-* request header, and a conforming client must drop a tool whose annotation is invalid from tools/list 3.1

tools/list, prompts/list, resources/list, resources/read and resources/templates/list results now require ttlMs and cacheScope through a CacheableResult interface; ttlMs is a freshness hint in milliseconds and cacheScope, either public or private, controls whether shared intermediaries may cache the response (SEP-2549) 1.2 Servers should also return tools from tools/list in a deterministic order, which the changelog ties to client-side caching and LLM prompt cache hit rates.1 Client and server capabilities also gain an extensions field, and the changelog documents OpenTelemetry trace context propagation for _meta keys such as traceparent, tracestate and baggage (SEP-414).1

One phrasing difference is worth knowing before you write compatibility code. The release post presents server/discover as an optional capability-discovery call that clients may use up front, while the changelog says servers must implement the RPC, which still leaves clients free not to call it 2.1

MRTR replaces the held-open-stream flows

Multi Round-Trip Requests (SEP-2322) replace the server-initiated requests that previously needed a constantly open stream. A tool that needs a confirmation or a missing parameter returns a result marked input_required with an inputRequests field naming what it needs; the client answers by retrying the original call with inputResponses attached 2.1 Every result now carries a required resultType, complete for ordinary results, and clients must treat results from earlier-protocol servers that omit the field as complete.1

The Inspector documentation shows what that means in practice: a tool can return input_required instead of a final result, embedding an elicitation, a sampling request or a roots/list request, and the client answers and retries the tools/call under a fresh JSON-RPC id until the call reaches complete.3 The older pattern of a server calling elicitation directly errors outright on a 2026-07-28 connection, because server-to-client requests are not allowed there.3

Deprecations and the twelve-month window

Roots, Sampling and Logging are deprecated (SEP-2577). They remain part of the specification and keep working through the deprecation window, but new implementations should not add support for them; the changelog suggests passing directories or files through tool parameters, resource URIs or server configuration instead of Roots, integrating directly with LLM provider APIs instead of Sampling, and logging to stderr or OpenTelemetry instead of Logging.1 The HTTP+SSE transport, described as deprecated since protocol version 2025-03-26, is formally reclassified as Deprecated with a migration path to Streamable HTTP (SEP-2596), and the includeContext values thisServer and allServers move from soft-deprecated to Deprecated.1

Behind all of this is a governance change: a feature lifecycle policy defining Active, Deprecated and Removed states, a minimum twelve-month deprecation window and a registry of deprecated features (SEP-2596) 1.2 The retained evidence does not state when that window starts for each feature, so it does not support computing removal dates for Roots, Sampling, Logging or the legacy transport.

Authorization hardening and the DCR-to-CIMD shift

Authorization is where the changelog adds requirements rather than suggestions. Authorization servers should include the iss parameter per RFC 9207, and MCP clients must validate a present iss against the recorded issuer before redeeming an authorization code (SEP-2468).1 Whether an absent iss is itself a failure is not settled by the two available phrasings, which differ on whether validation is framed as unconditional or limited to a present parameter 2.1

Client credentials are bound to the authorization server that issued them: clients must key persisted credentials by issuer, must not reuse them with a different authorization server, and must re-register when that server changes (SEP-2352).1 Clients must also set an appropriate application_type during Dynamic Client Registration, which the changelog connects to OpenID Connect redirect URI conflicts that have rejected localhost redirects for desktop and CLI apps (SEP-837) 1.2

Dynamic Client Registration itself is now formally deprecated in favour of Client ID Metadata Documents. DCR keeps working for backward compatibility with authorization servers that do not support CIMD, and the changelog says it will be removed in a future version of the spec, so this is a future removal rather than a current break.1

SDK boundaries: TypeScript and Python verified, others qualified

The two SDK sources here are registry metadata rather than migration guides, and they draw different lines.

  • TypeScript: @modelcontextprotocol/sdk is a v1.x maintenance line at 1.32.2 that implements the spec up to 2025-11-25, and 2026-07-28 support is explicitly not planned for v1.x. Support ships in v2 as separate packages, @modelcontextprotocol/server and @modelcontextprotocol/client, and v1.x continues to receive bug fixes and security updates for at least six months after the v2 release on 2026-07-27.4
  • Python: the mcp package is at 2.3.0 with a reworked surface, an MCPServer class and a Client class, and the project publishes its migration guide separately.5

The release post states that all four Tier 1 SDKs (TypeScript, Python, Go and C#) speak 2026-07-28 on the day of publication and that the Rust SDK supports the new spec in beta.2 No version numbers for Go, C# or Rust appear in the evidence here, and neither the TypeScript nor the Python migration guide body was captured, so step-level instructions for removing Mcp-Session-Id from a specific codebase are not verified by these sources.

Debugging the cutover with the Inspector's protocol eras

The Inspector treats protocol era as a per-server setting orthogonal to transport, so the same HTTP URL can be inspected as a legacy or a modern server.3 Three modes exist: legacy performs a plain initialize with no probing and is the default; auto probes server/discover first and falls back to initialize on any non-modern outcome; modern pins 2026-07-28 exactly with no fallback, so a non-modern server fails loudly rather than degrading silently.3 The documentation explains that the default avoids auto-probing on purpose, because a probe stalls against silent legacy stdio servers and pollutes the transcript you are trying to read.3

Two behaviours are worth wiring into your own test setup. On a modern connection, log level is stamped per request, and omitting it means the same tool call produces no logs at all, which the documentation calls correct rather than a bug.3 And because subscriptions become a long-lived stream rather than a session flag, the Inspector exposes stream states such as Connecting, Listening, Reconnecting, Stream ended and Not acknowledged, and it does not retry when a server answers a listen without acknowledging it.3

When migration is worth it

The sourcing here supports an operational argument more than a performance one. Statelessness takes shared storage out of the request path, makes header-level routing and metering possible, and lets gateways cache list results with ttlMs and cacheScope 1.2 The maintainers describe the migration cost as concentrated on developers who depended on session identifiers, and say early testing feedback made the process easier.2

Everything beyond that in the release post is a vendor statement, not a measurement. Cloudflare says its Agents SDK supports the spec from day zero; AWS says the stateless core is available in Amazon Bedrock AgentCore and that it contributed the Tasks extension; Manufact reports roughly 83% smaller packages and 25% faster performance with SDK v2 and its client-server split; Supabase describes MRTR as the reason it can now support elicitation despite running statelessly.2 Those are endorsements from partners who helped validate the spec, and the retained evidence contains no benchmark, reliability or cost numbers to check them against. Supabase's statement is an intention to support elicitation, not confirmation of shipped support.

The practical decision therefore rests on your own sessions and SDK line. If your server keeps state in the transport, that is a real code change: mint handles, move log level to per-request, and re-issue requests after a broken stream.1 If you are on TypeScript v1.x, the boundary is sharper, because 2026-07-28 support is not planned for that line and the v2 packages are the path.4 Against that urgency, Roots, Sampling, Logging and the legacy HTTP+SSE transport keep working for at least twelve months, so a client that only consumes those is not forced to move today 2.1 One open point worth checking against your own design: the evidence does not state whether client-initiated tasks survive the redesign, and in the modern era task creation is described as server-directed.3

Sources

  1. 1.modelcontextprotocol.io (modelcontextprotocol.io)
  2. 2.blog.modelcontextprotocol.io (blog.modelcontextprotocol.io)
  3. 3.modelcontextprotocol.io (modelcontextprotocol.io)
  4. 4.registry.npmjs.org (registry.npmjs.org)
  5. 5.pypi.org (pypi.org)