The Standard Everyone Adopted, Fast

By April 2026, 17,468 MCP servers were indexed across public registries. The Model Context Protocol had accumulated 97 million SDK installs. In less than 18 months, it went from an Anthropic internal project to the closest thing agentic AI has to a universal plumbing standard. That speed is genuinely impressive — and it is exactly why what happens to MCP’s governance now matters more than its technical design.

On December 9, 2025, Anthropic donated MCP to the newly formed Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation. The press release used all the right words: open, neutral, collaborative, community-driven. OpenAI and Block joined as co-founders. Google, Microsoft, AWS, Cloudflare, and Bloomberg signed on as Platinum members. By February 2026, 97 more companies had joined. The message was clear: MCP is no longer Anthropic’s protocol. It belongs to the industry.

That framing is worth examining more carefully.

What “Platinum Member” Actually Means

The AAIF’s governing board makes decisions on strategic investments, budget allocation, membership approvals, and overall direction. Each Platinum member — AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI — gets to appoint a representative to that board. This is standard Linux Foundation practice, and it is not inherently problematic. The Linux kernel thrived under a similar model. Kubernetes did too.

But the Linux kernel is a collaborative infrastructure project whose major contributors are not also its most intense market competitors. MCP’s governing board looks different. OpenAI and Anthropic are direct rivals in the frontier model market. Google and Microsoft are competing for enterprise AI contracts worth hundreds of billions of dollars. AWS competes with all of them on inference infrastructure. These companies are not just co-governing a shared resource — they are each trying to win a market that MCP’s design choices will directly affect.

The Linux Foundation provides genuine neutrality for the administrative layer: legal home, trademark enforcement, infrastructure. But neutrality on procedural questions does not resolve disagreement on technical ones. When OpenAI’s interests in a particular MCP extension diverge from Google’s, the governing board does not automatically produce a neutral answer. It produces a negotiation — and negotiations between well-resourced incumbents tend to favor whoever is most willing to invest engineering time in the outcome.

The Fragmentation Risk Is Already Visible

The 2026 MCP roadmap correctly identifies governance maturation as a priority. Working Groups, Specification Enhancement Proposals, a contributor ladder — all of this is genuine progress. But the risk of an “open” standard governed by competing vendors is not capture by a single bad actor. It is fragmentation: a stable core spec that everyone nominally supports, surrounded by a proliferating set of proprietary extensions that only work well within one vendor’s ecosystem.

This is exactly what happened to HTML in the 2000s. Internet Explorer 6 was nominally compliant with open web standards. Microsoft also shipped dozens of IE-specific extensions — ActiveX, VBScript, proprietary CSS filters — that made the web work better on IE and worse everywhere else. Value migrated to the proprietary layer, not the open standard. The WHATWG was formed in 2004 precisely because the W3C’s multi-vendor governance was too slow and too dominated by incumbent interests to keep up.

The same dynamic is already appearing in MCP. Multiple vendors are shipping platform-specific MCP extensions — capabilities baked into their SDKs, their hosting environments, their authentication flows — that work correctly only within their stacks. The core MCP spec remains portable. The useful things built on top of it, increasingly, are not. By the time a user has invested in an MCP server architecture tightly integrated with a major cloud platform, they are not really running a neutral open standard. They are running that platform’s interpretation of one.

Google Has a Backup Plan

One structural signal worth noting: Google is both a Platinum member of the AAIF and the primary backer of Agent-to-Agent (A2A), an alternative agentic communication protocol it has been developing in parallel. A2A and MCP serve overlapping use cases — connecting AI systems to tools and to each other. Google publicly frames them as complementary. That may be true today. It is also true that Google is not betting its agentic AI strategy exclusively on a protocol where Anthropic has more institutional knowledge, a head start in developer mindshare, and formal governance influence equal to Google’s own.

A company with a seat on MCP’s governing board and an active competing standard is not the same as a company that has made an unconditional commitment to MCP. That is not a criticism of Google — hedging is rational when the stakes are this high. But it should inform how developers think about the long-term stability of MCP as a standard versus MCP as a convention. Conventions can bifurcate. Standards are supposed to prevent that.

We explored MCP’s explosive growth earlier this year in MCP at 9,400 Servers: Agentic Infrastructure Outruns the Models. The adoption numbers have only accelerated since. What has not accelerated at the same pace is the governance machinery needed to manage what happens when these platinum-member interests collide.

“Open” Is Still Better Than the Alternative

None of this is an argument that MCP’s Linux Foundation move was a mistake. The alternative — a single-vendor proprietary standard controlled by Anthropic — would be strictly worse. Linux Foundation governance means the spec is publicly available, changes are documented, and no single company can unilaterally break compatibility without visible accountability. That matters.

But “open” in AI infrastructure often means something narrower than developers assume: open for inspection, open to comment, open for membership if you can afford Platinum dues. It does not automatically mean governed in the interest of the smallest server operator or the startup building on top of the protocol. The companies that show up to working group meetings, file the most SEPs, and employ the most active maintainers will shape what MCP becomes. Those companies are disproportionately the same ones with governing board seats.

Developers building production agentic systems on MCP today should make two assumptions: the core transport and tool-call semantics will remain stable, because all eight Platinum members need that stability. The higher-level patterns — auth, multi-agent orchestration, registry discovery — will evolve in ways that reflect the commercial interests of the companies investing most heavily in the working groups. Knowing which is which, and building accordingly, is the practical skill that MCP’s rapid adoption has made necessary.

Further Reading

• Linux Foundation: AAIF Announcement — The founding press release, including the full list of Platinum members and the governance structure of the new foundation.
• The 2026 MCP Roadmap — The official priorities for the year, including governance maturation, transport scalability, and enterprise readiness.
• Browser Dominance, Standards Setting, and WHATWG vs W3C — A sharp historical analysis of how browser vendor interests shaped the HTML standard, with lessons that apply directly to AI protocol governance.