MCP Goes Stateless: What the July 28 Spec Means for Your AI Agents

30-Second Summary

The Model Context Protocol is shipping its biggest revision since launch on July 28, 2026. The headline change: MCP becomes stateless. Sessions, the initialize handshake, and the Mcp-Session-Id header are gone. Any MCP server that previously required sticky sessions and a shared session store can now run behind a standard load balancer. Four Tier 1 SDK beta releases are available today. Operators running or building AI agents have eight days to review the changes before the spec becomes official.

---

At a Glance

• Topic: AI Infrastructure
• Company: Model Context Protocol / Anthropic
• Date: July 20, 2026 (final spec ships July 28, 2026)
• Announcement: Release candidate for MCP specification 2026-07-28, the largest protocol revision since launch
• What Changed: Sessions removed; protocol becomes stateless; MCP Apps and Tasks extension added; OAuth/OIDC security hardened; W3C distributed tracing formalised
• Why It Matters: Dramatically simplifies enterprise deployment of AI agents at scale, aligns with enterprise security standards, and adds UI capabilities without additional infrastructure
• Who Should Care: Engineering leads and CTOs building AI agents, security teams responsible for AI infrastructure, operators evaluating whether to standardise on MCP for their agent stack

---

Key Facts

• The release candidate has been locked since May 21, 2026. The final spec publishes July 28, 2026.
• The initialize handshake and Mcp-Session-Id header are removed from the Streamable HTTP transport.
• Requests now carry Mcp-Method and Mcp-Name headers, enabling load balancers to route traffic without inspecting message bodies.
• List endpoints (tools/list, resources/list, prompts/list) now return ttlMs and cacheScope parameters, allowing clients to cache results like HTTP Cache-Control directives.
• Beta SDKs are available: Python v2.0.0b1, TypeScript v2 beta, Go v1.7.0-pre.1, C# v2.0.0-preview.1.
• Three features enter the 12-month deprecation pipeline: Roots, Sampling, and Logging.
• Breaking change: the 2025-11-25 Tasks API is replaced by the new Tasks extension; error code -32002 updates to -32602.
• "Nothing breaks on July 28" for implementations not using deprecated features. The date marks specification publication, not forced migration.

---

What Happened

The Model Context Protocol, which standardises how AI models connect to external data sources and tools, has been in active production use since 2025. The 2026-07-28 specification is the first major architectural revision and the result of a public RFC process that began earlier this year.

The core change removes the stateful session layer entirely. Previously, an MCP server required a session store shared across instances, sticky routing at the load balancer, and a protocol handshake before any work could begin. The new architecture pushes client metadata into the _meta field on every request. A server that once needed specialised gateway infrastructure can now run as a stateless HTTP service on any cloud platform's standard container runtime.

Alongside the stateless core, two experimental features have been graduated to official extensions. MCP Apps allows servers to ship interactive HTML interfaces rendered in sandboxed iframes, with UI interactions flowing through the same JSON-RPC protocol as direct tool calls. This means an agent can surface a form, a dashboard, or a review interface inside any MCP-capable client without a separate frontend deployment. The Tasks extension formalises the model for long-running work: servers can return task handles from tool calls, and clients drive progress through tasks/get, tasks/update, and tasks/cancel.

Six changes to the authorisation layer align MCP more closely with enterprise security requirements. Clients must now validate issuer parameters under RFC 9207, declare application type during OAuth registration, and bind credentials to specific authorisation servers. W3C Trace Context propagation is also formally documented, enabling distributed tracing across SDKs and gateways using standard OpenTelemetry backends.

---

Why It Matters

Simpler enterprise deployments at scale. The sticky session requirement was a genuine infrastructure tax. Removing it means MCP servers can run on any standard container platform, auto-scale without state synchronisation, and sit behind a commodity load balancer. This reduces the operational overhead of running agent infrastructure by a meaningful margin.

Security alignment is overdue. Enterprise security teams have been cautious about MCP in part because the OAuth implementation diverged from standard patterns. The six hardening changes in this spec, particularly the RFC 9207 issuer validation and credential binding, bring MCP into alignment with the patterns security teams already review and approve in enterprise software procurement.

MCP Apps changes the interface question. One of the practical limitations of agent deployments has been that agent outputs are text-based, requiring separate UI work to surface structured interactions. MCP Apps allows agents to deliver interactive interfaces inside the client itself. For operators building internal tools on top of AI agents, this removes a layer of frontend development from the delivery path.

A deprecation policy creates predictability. The introduction of a formal 12-month deprecation policy, covering Roots, Sampling, and Logging, is as significant as the features themselves. Enterprise adoption of any standard depends on confidence that the protocol will evolve predictably. A published deprecation timeline signals that MCP is maturing from a fast-moving specification into an infrastructure standard.

Eight days to plan. The final spec ships July 28. Teams that have invested in MCP-based agent infrastructure have a short window to review breaking changes, test beta SDKs, and schedule any migration work. This is not a forced cutover, but early testing against real workloads is the advised approach.

The ecosystem consequence. When Tier 1 SDKs (the languages most enterprise AI agent tooling is built in) ship version 2 betas simultaneously with the specification release candidate, it signals coordinated ecosystem readiness. The gap between specification and production-grade SDK support, which historically delayed enterprise adoption of new protocols, is shorter here than in most comparable transitions.

---

The David and Goliath View

MCP's stateless transition is the moment the protocol stops being infrastructure for early adopters and starts being infrastructure for enterprise IT. Stateless HTTP is not a novel concept. The significance here is that MCP has been holding back adoption by requiring deployment patterns that sit outside most enterprise organisations' standard operating procedures. Removing sessions removes the main objection infrastructure and security teams had when evaluating MCP for production use.

For operators running 10 to 200 person companies, the more immediate implication is competitive. The tooling around AI agents is standardising faster than most businesses are moving. MCP is becoming the plumbing layer for connecting AI to internal systems, and the July 28 specification is the version that will define production deployments for the next two to three years. Organisations that understand this version and have hands-on experience with it will build faster and with more confidence than those starting from scratch.

The MCP Apps extension is worth watching closely, separate from the stateless headline. The ability to ship interactive interfaces through the same protocol as tool calls is a significant simplification for anyone building internal AI-powered workflows. It may not matter today, but it quietly removes a constraint that has been quietly shaping what operators think is possible with AI agents.

---

Where This Fits in the AI Stack

MCP sits at the connection layer between AI models and the systems those models need to act on. In the typical enterprise AI architecture, a model like Claude or GPT-5.6 connects to databases, APIs, calendars, CRMs, and internal tools through an integration layer. MCP is the specification that standardises how those connections work.

The 2026-07-28 update affects the plumbing, not the model. It makes the plumbing cheaper to run, easier to secure, and more capable of surfacing results. For operators focused on business outcomes rather than infrastructure, the relevant implication is that building reliable agent-to-tool connections is now less technically demanding, which means the constraint on what AI agents can do inside a business is shifting from infrastructure to imagination.

---

Questions Operators Are Asking

Do we need to update our existing MCP servers immediately? No. The specification description is explicit: "nothing breaks on July 28." The date marks publication, not forced migration. However, if you are using the Tasks API from the 2025-11-25 spec, that is a breaking change and migration is required before the deprecation window closes. For all other implementations, plan migration into your next infrastructure cycle.

What does stateless mean for our security review process? It simplifies it. The previous session architecture required custom review of the session management implementation. A stateless architecture follows the same patterns as any standard HTTP API. Combined with the OAuth/OIDC hardening, security teams familiar with reviewing REST APIs now have a clear framework to assess MCP deployments.

What should our engineering team do this week? Install the beta SDK for your primary language and run it against a real production workload or staging environment. The specification recommends testing against live traffic rather than basic test cases, specifically for HTTP deployments and stateless path compatibility. Pin the exact beta version to avoid unexpected API changes during the test period.

Is MCP worth investing in if we haven't started yet? The specification reaching this level of maturity, with a formal deprecation policy, Tier 1 SDK parity, and broad enterprise adoption, is the signal that investment now compounds rather than risks. Waiting for a protocol to stabilise is a reasonable strategy. For MCP, the wait is effectively over with this release.

What happened to the Roots, Sampling, and Logging features we use? All three enter deprecation under a 12-month minimum window, meaning they will not be removed until at least mid-2027. The official replacements are tool parameters or resource URIs for Roots, direct LLM provider API integration for Sampling, and stderr or OpenTelemetry for Logging. Plan migration, but there is no urgency.

---

Citable Summary

The Model Context Protocol specification 2026-07-28 ships on July 28, 2026 and represents the largest revision to the protocol since its launch. The core change removes stateful sessions, allowing MCP servers to run on standard HTTP infrastructure without sticky routing or shared session stores. New additions include MCP Apps for server-rendered interactive interfaces, a formally graduated Tasks extension for long-running work, and six OAuth/OIDC authentication hardening changes. Beta SDKs are available today in Python, TypeScript, Go, and C#. Three features (Roots, Sampling, Logging) enter a 12-month deprecation pipeline. For operators and engineering teams: nothing breaks on July 28, but teams using the 2025-11-25 Tasks API must migrate. The specification is available at blog.modelcontextprotocol.io.