MCP governance in 2026: what the Linux Foundation handoff changed

Anthropic donated MCP to the Linux Foundation in December 2025. Who decides the spec now, what the 2026-07-28 release removed, and what to check.

AI and AutomationAugust 21, 20268 min read

MCP governance is the set of rules that decide who can change the Model Context Protocol, how a change gets approved, and when it ships. Since December 2025 those rules sit with the Agentic AI Foundation, a directed fund inside the Linux Foundation, and no longer with Anthropic alone.

This matters to any team running an MCP server in production or budgeting one for next year. A protocol owned by one vendor moves on that vendor’s schedule and for that vendor’s reasons. A protocol under a foundation moves through a written proposal process, with a public trail and a maintainer group drawn from companies that compete with each other. The first specification produced end to end under the new arrangement, dated 2026-07-28, is also the most disruptive release MCP has shipped. If you are new to the protocol itself, start with what MCP is and why a SaaS needs a server.

What happened in December 2025

On 9 December 2025 Anthropic donated MCP to the newly formed Agentic AI Foundation (AAIF), created as a directed fund under the Linux Foundation. AAIF opened with three founding projects: MCP from Anthropic, AGENTS.md from OpenAI, and the goose agent framework from Block. The Linux Foundation announcement named Google, Microsoft, AWS, Cloudflare and Bloomberg among the backers.

The donation followed adoption rather than creating it. At the handoff, MCP reported more than 97 million monthly SDK downloads and around 10,000 active servers, with first-class client support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. A protocol at that scale living inside one company is a procurement objection with a long shelf life. Moving it to a neutral home answers the objection without changing a line of the spec.

Who decides what goes into the spec now

Two bodies, two jobs. The AAIF governing board handles strategy, budget, member recruitment and approval of new projects. Technical direction stays with the MCP maintainers. The project put the split plainly at the time of the move: the people making decisions about the protocol are still the maintainers who were already stewarding it.

Changes arrive as SEPs, Specification Enhancement Proposals. An SEP is a written proposal with an owner, a review cycle and a public record. Working groups cover domains such as transports, agents and governance itself. The maintainer group grows in the open: an April 2026 update added Clare Liguori as a core maintainer and Den Delimarsky as a lead maintainer.

Two items on the 2026 roadmap matter more than the org chart. The first is a contributor ladder: a documented path from community participant to working group contributor, facilitator, lead maintainer and core maintainer, with named criteria at each step. The second is delegation. Trusted working groups can accept SEPs in their own domain without waiting for a full core review, so core maintainers keep strategic direction while working groups get execution speed.

There is also a gate that did not exist before. A Standards Track SEP cannot reach Final status until a matching scenario lands in the conformance suite. Said in plain terms: a feature is not final until a test exists that proves an implementation supports it. Conformance test suites and SDK tiers are funded as continuous work under SEP-1730, not as a one-off cleanup.

What a platinum seat buys, and what it does not

The foundation has grown fast. In its membership announcement AAIF reported 146 members after adding 18 Gold and 79 Silver members, and appointed David Nalley, director of developer experience at AWS, as governing board chair. The platinum tier holds eight companies: AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.

None of that converts into merge rights on the specification. A board seat decides where the foundation spends money and which projects it accepts. Maintainers decide what the protocol says. Keeping those two levers apart is the reason a directed fund exists, and it is the thing to verify the next time a vendor describes its own MCP extension as standard. Ask which SEP it came from and what status that SEP holds.

The 2026-07-28 spec: the first release written under the new process

The release candidate locked on 21 May 2026 and the final specification published on 28 July 2026. It is the largest revision since MCP launched, and it lands in three parts.

The core went stateless

The initialize and initialized handshake is gone, along with the protocol-level session and the Mcp-Session-Id header. Removed, not deprecated. Client information, protocol version and capabilities now travel in the _meta field of every request, so each call describes itself. A remote server can sit behind an ordinary round-robin load balancer with any instance answering any request, no session affinity and no shared session store.

This is the change that costs engineering time. Older clients keep working only if you run the old path alongside the new one, which means two code paths and a date on which you turn the first one off.

Extensions became a governed track

Extensions used to be an informal convention. They are now a framework, with an Extensions Track inside the SEP process that defines how an extension graduates from experimental to official. The first official set includes MCP Apps for sandboxed HTML interfaces, a stateless version of Tasks, Enterprise Managed Authorization and OAuth Client Credentials.

Authorization moved toward ordinary OAuth

Core authorization now tracks production OAuth 2.0 and OpenID Connect deployments more closely: issuer validation, client credentials bound to their authorization server, and a shift away from Dynamic Client Registration toward Client ID Metadata Documents. The Enterprise Managed Authorization extension lets an organization run MCP access through an identity provider it already owns, such as Okta or Microsoft Entra ID. We covered the operational side of that in hardening an MCP server for enterprise.

What to check if you already run an MCP server

Grep for session state. Every read of Mcp-Session-Id, every assumption that a handshake ran before the first tool call, every in-memory map keyed by session. Those are your migration surface.
Pick a dual-path window and put a date on it. Running old and new transports side by side is supported. Running them forever is how the second path never ships.
Move per-session state to a store or delete it. Stateless routing means the instance that answers a request may never have seen the client before.
Re-read your auth against the new rules. If you rely on Dynamic Client Registration, plan the move to Client ID Metadata Documents and add issuer validation.
Track the conformance suite, not the blog posts. When a feature you depend on has a conformance scenario, you have something testable to point at during a security review.

Client support arrives on each vendor calendar. Anthropic published its own rollout plan for Claude separately from the spec date, and other platforms did the same.

What foundation governance does not fix

Three limits worth naming.

Governance sets who decides, not how fast. A process with working groups, review cycles and a conformance gate is slower than one company shipping a change on a Tuesday. Stability is what you buy with that time.

The spec is not the implementations. A feature reaching Final status tells you the standard settled. It tells you nothing about whether the three clients your customers use have shipped it.

Extensions can still fragment the field. A governed track makes fragmentation visible and reviewable. It does not make it impossible. If a vendor extension is the only reason your server works with that vendor’s client, you hold a dependency no foundation can remove for you.

For a team weighing whether to run its own server or buy one, the handoff tilts the calculation slightly toward building. The spec is now harder for any single vendor to move under you, and the proposal trail gives you months of warning before a breaking change lands. The rest of that decision, cost and maintenance load included, sits in managed versus self-hosted MCP servers.

Sources

In this piece

Photo by Phyllis Lilienthal ↗ on Unsplash ↗

Frequently asked questions

Does the Linux Foundation now control MCP?+

No. The Linux Foundation hosts the Agentic AI Foundation, which holds the assets and runs the budget. Technical decisions stay with the MCP maintainers through the SEP process. The governing board approves projects and spending, not merges on the specification. This is the standard split for a directed fund, and it is why a platinum membership does not buy influence over what the protocol says.

Do I have to migrate to the 2026-07-28 spec right away?+

Not immediately, but put a date on it. The initialize handshake and the Mcp-Session-Id header were removed rather than deprecated, so old clients only keep working while you run both paths side by side. Two paths mean two sets of bugs and two sets of tests. Most teams should plan the cutover for the quarter after their main client platforms ship support, and stop maintaining the old path once traffic on it drops below a threshold they set in advance.

How do I tell whether a vendor MCP extension is actually standard?+

Ask for the SEP number and its status. Extensions now graduate through an Extensions Track inside the SEP process, so an official extension has a public proposal, a review trail and a status you can read. If a vendor cannot name the SEP, the extension is theirs, not the protocol’s. That is allowed and sometimes useful, but it is a dependency on one vendor and it belongs in your risk register, not in your architecture diagram as a standard.

Does neutral governance make MCP safe to standardize on for the next five years?+

It removes one risk and leaves others. Vendor capture is much harder now: the spec has a written process, a cross-company maintainer group and a conformance gate before a feature reaches Final. What governance cannot promise is that MCP wins the category. Agent interoperability is still being contested, and AAIF itself hosts competing pieces such as AGENTS.md and goose. A five-year bet is reasonable for the tool-calling surface of a product, and less reasonable for anything you cannot re-implement behind an internal interface.

Related services

Start a project.

We write about what we build. Tell us what you want to build.

Talk to the studio Read more articles