The State of MCP, 2026

The short version: MCP changed owners and changed shape in the same eight months. Governance moved from a single vendor to a multi-company foundation in December 2025; the protocol itself then went through its most disruptive technical redesign since launch in July 2026, dropping stateful sessions entirely. Anything integrated against a pre-2026-07-28 spec revision needs an active migration plan, not a set-and-forget integration — and the governance change means future revisions like it will keep coming through an open process, not a single company's roadmap.

• Anthropic donated MCP to the Agentic AI Foundation on December 9, 2025, alongside two other founding projects — Block's goose and OpenAI's AGENTS.md — it's no longer a single-vendor format ^[1].
• The 2026-07-28 spec revision dropped stateful sessions entirely in favor of a stateless request model, and moved routing into dedicated HTTP headers — MCP's biggest architectural change to date ^{[3] [4]}.
• That statelessness shift created a new named attack surface: state handle hijacking.
• Self-reported adoption: 97M+ monthly SDK downloads, 10,000+ active servers at donation time — third-party registries report 71,000+ by August 2026, uncorroborated and likely not deduplicated ^[6].

Governance: From Single-Vendor to Foundation

The strongest signal about where MCP stands is governance, not technology. On December 9, 2025, Anthropic donated MCP — which it had released as its own open standard in November 2024 — to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI ^[1]. MCP joined as one of three founding projects, alongside Block's goose agent framework and OpenAI's AGENTS.md, the de facto format for giving coding agents build steps and guidelines ^[2]. Amazon Web Services, Google, Microsoft, Bloomberg, and Cloudflare joined as additional platinum members.

Anthropic's own stated rationale was that donation "ensures MCP remains open and community-driven" under the same neutral stewardship model that supports Kubernetes, PyTorch, and Node.js. The practical effect: MCP is no longer a single-vendor integration format to evaluate for lock-in risk — it's closer to infrastructure, which changes the calculus for building on it. It also means the protocol now moves through open Working Groups rather than one company's release cycle, which is the direct cause of the disruptive July 2026 redesign below — a foundation-run process optimizing for the ecosystem's stated priorities, not a single vendor's ship schedule.

The 2026-07-28 Redesign: Statelessness and Gateway-Native Routing

The corresponding technical takeaway is that the protocol is still actively and substantially changing under that new governance. The 2026-07-28 spec revision is the single largest architectural change in MCP's history: it dropped the stateful initialize handshake and the Mcp-Session-Id header entirely, in favor of a fully stateless model where every request carries its own version and capability metadata ^[3].

The less-discussed but equally consequential change is where routing information now lives. Method and tool names moved out of the JSON-RPC body and into dedicated Mcp-Method and Mcp-Name HTTP headers — which means a gateway, rate limiter, or WAF sitting in front of an MCP server can now route, meter, and authorize requests by reading headers alone, without parsing or understanding the JSON-RPC payload itself ^[4]. Authorization was hardened to match current OAuth 2.0 and OpenID Connect practice, add issuer validation, and bind client credentials to their authorization server — with the spec beginning a deliberate move away from Dynamic Client Registration toward Client ID Metadata Documents ^[4]. tools/list, resources/list, and prompts/list no longer vary per-connection and are now cacheable. Any server that needs cross-call state must mint an explicit handle passed back as an ordinary argument on every subsequent call, rather than relying on the protocol's implicit session identity.

• Removed: the initialize/notifications/initialized handshake, session headers, SSE stream resumability and the Last-Event-ID redelivery mechanism.
• Added: Multi Round-Trip Requests (a synchronous input_required/inputResponses pattern replacing bidirectional streams), header-based method/tool routing, a mandatory server/discover RPC, a formal extensions framework, and an Active/Deprecated/Removed feature lifecycle with a twelve-month minimum deprecation window.
• Consequence: a new named attack surface — state handle hijacking — since state continuity is now the server's explicit responsibility rather than the protocol's. Covered in MCP Security in Production.

Adoption

Third-party adoption figures, self-reported by Anthropic and AAIF at the time of the foundation donation: over 97 million monthly SDK downloads and 10,000+ active servers, with first-class client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and VS Code, plus Block's goose agent framework^[1].

The ecosystem's own registries report far higher and inconsistent counts. Glama's registry alone grew from roughly 100 servers at MCP's November 2024 launch to over 19,800 by March 2026, then past 71,000 by August 2026; the MCP Toplist separately tallies nearly 101,000 combined across registries ^[6]. But registry-listed does not mean maintained or secure — a distinction that matters more given the documented supply-chain and configuration-trust incidents covered in MCP Security in Production.

The 2026 Roadmap

Rather than commit to a dated release schedule, MCP's maintainers published their 2026 roadmap organized around four priority areas instead — an explicit acknowledgment, in the maintainers' own words, of "the uncertainty inherent in a fast-growing project like MCP" ^[5]: Transport Evolution (continuing to harden the stateless, header-routable transport introduced in 2026-07-28), Agent Communication (closing the gap between MCP's single-agent-to-tools model and multi-agent coordination, where A2A currently sits), Governance Maturation (deepening the Working Group process now that AAIF, not Anthropic, owns the roadmap), and Enterprise Readiness (the audit, compliance, and operational tooling gap that shows up directly in the governance frameworks covered below).

Where This Fits

• What MCP and its complementary protocol A2A actually specify at the mechanics level, and when each is warranted: MCP and A2A in Production and MCP vs A2A.
• The broader agentic AI production landscape MCP sits inside, including the governance frameworks racing to catch up with agent-specific risk: The State of Agentic AI in Production, 2026.