Standards
MCP's biggest revision: what the 2026-07-28 spec changes
The 2026-07-28 Model Context Protocol specification - billed by its maintainers as the largest revision of the protocol since launch - replaces MCP's session-based handshake with a stateless core, adds two official extensions (MCP Apps and Tasks), and deprecates three long-standing features: Roots, Sampling and Logging. Agent builders should care, because the changes touch how clients connect, authorise and carry per-request metadata - so anything built against the old session model needs revisiting rather than a drop-in upgrade.
Dated analysis. Checked 18 July 2026 against the release candidate, ten days before the specification's final publication on 28 July 2026. The RC validation window ran from 21 May to 28 July 2026, so the core shape below is settled, but exact wording could shift before the final cut. Adoption figures are third-party survey aggregations, not a single primary dataset, and are reported with their sources attached.
What changes in the core?
The headline change is the move to a stateless core. The
revision removes the session handshake and the Mcp-Session-Id
model, so an MCP server no longer has to hold per-connection session state and
can scale on ordinary HTTP infrastructure. Client metadata that used to ride
the session now moves into per-request _meta fields. Alongside
that, the spec introduces a formal Extensions framework with
reverse-DNS identifiers and independent versioning and governance, so
capabilities can evolve outside the core without forcing a whole-protocol bump.
What are the new extensions?
Two official extensions ship with the revision. MCP Apps defines server-rendered, sandboxed-iframe UIs - a way for a server to present interface elements rather than only tool results. Tasks graduates from experimental to a full extension, with an explicit task-handle lifecycle for longer-running or asynchronous work. Both live in the new extensions framework rather than the core, which is the point: they can move at their own pace.
| Area | Change |
|---|---|
| Core transport | Stateless; session handshake removed |
| Client metadata | Moves to per-request _meta |
| New extension | MCP Apps (sandboxed-iframe UIs) |
| New extension | Tasks (experimental → extension) |
| Authorization | Six OAuth 2.0 / OIDC-aligned proposals |
| Deprecated | Roots, Sampling, Logging |
What about authorisation?
Authorisation gets the most concerted attention. The revision carries six enhancement proposals aligning MCP closer to OAuth 2.0 and OpenID Connect, including RFC 9207 issuer validation. Separately, the Enterprise-Managed Authorization extension has been promoted to stable, giving organisations centralised, identity-provider-based access control over MCP servers; InfoQ reports it has been adopted by Anthropic, Microsoft and Okta among others. For anyone deploying MCP inside an enterprise, this is the change with the most immediate operational weight.
What gets deprecated?
Three features - Roots, Sampling and Logging - are marked deprecated, each with a specified replacement. Deprecation is not immediate removal, but it is the signal to stop building new work against them. If your agent leans on any of the three, the migration path is the thing to read closely in the final spec.
How widely is MCP actually used?
Here the numbers deserve scepticism, because they diverge sharply. Reported SDK activity is large and consistent - roughly 97 million monthly MCP SDK downloads as of July 2026. Adoption rates are where sources disagree: one tracker cites 78% of enterprise AI teams running MCP-backed agents in production and 28% of Fortune 500 companies operating MCP servers, while a more conservative reading - Stacklok's 2026 software report - puts 41% of surveyed software organisations in limited-or-broad MCP production use. Those figures come from different survey panels and methodologies, so the honest statement is a range: somewhere between roughly 41% and 78% of the surveyed populations, depending on who was asked and how "in production" was defined. Anyone quoting a single blended adoption number is smoothing over a genuine disagreement between the underlying surveys.
The practical takeaway for builders is that the protocol underneath the tools is moving, and moving in a direction - stateless, extension-based, OAuth-aligned - that rewards designs which do not assume a persistent session. As with any capability change, the way to know whether the new shape helps your system is to prove it against your own workload, not the announcement.
Questions people ask
What's new in the MCP 2026-07-28 spec? A stateless core that drops the session handshake, two new official extensions (MCP Apps and Tasks), six OAuth 2.0 / OIDC-aligned authorisation proposals, and the deprecation of Roots, Sampling and Logging - billed by the maintainers as the largest revision since launch.
What happened to MCP sessions in the new spec? The
session-based handshake and Mcp-Session-Id model are removed in
favour of a stateless core; client metadata moves to per-request
_meta fields, so servers can scale on ordinary HTTP infrastructure.
What is MCP Enterprise-Managed Authorization? An extension, now promoted to stable, that gives organisations centralised, identity-provider-based access control over MCP servers. InfoQ reports adoption by Anthropic, Microsoft and Okta among others.
How many companies are using MCP in production? It depends on the survey. One tracker reports 78% of enterprise AI teams and 28% of Fortune 500 companies; Stacklok's 2026 report puts 41% of surveyed software organisations in production use. The figures diverge by methodology, so treat the range - not a single number - as the answer.
Should agent builders care? Yes, if you have built against the session model, use Roots/Sampling/Logging, or run MCP in an enterprise where centralised auth matters - all three are directly affected. If your agent is simple and stateless already, the migration is lighter, but the deprecations still set the direction of travel.
Sources & caveats. Core changes, extensions, deprecations and the "largest revision since launch" framing come from the official MCP 2026-07-28 release-candidate post; the Enterprise-Managed Authorization adoption detail from InfoQ. Adoption percentages are third-party survey aggregations that disagree by a wide margin (78% vs 41%) and are reported as a range with each figure attributed, not as a single fact. Read against the release candidate on 18 July 2026, ahead of final publication on 28 July.