MCP Went From Anthropic Side Project to Linux Foundation Standard

Most companies that build a good developer tool spend years trying to keep it. Anthropic built one and then, thirteen months later, handed the keys to someone else. In December 2025, MCP — the Model Context Protocol — was donated to the newly formed Agentic AI Foundation under the Linux Foundation. If you were watching from outside, that looked like a strange thing to give up. If you were watching the protocol wars of the last three years, it looked like the only move that could have worked.

MCP shipped in November 2024 as an Anthropic spec: a standard way for an AI model to discover and call tools, read resources, and use prompts exposed by an external server, instead of every vendor inventing its own bespoke integration format. The pitch was simple — write your integration once, and any MCP-compatible client can use it. What's easy to forget now is how unproven that pitch was at launch. Plenty of protocols ship with a clean spec and a compelling README and never make it past the vendor that wrote them.

The problem with a protocol owned by one company

A protocol's value is a network effect, not a feature list. The more clients and servers speak it, the more valuable it is to speak it too — but that only works if the parties adopting it trust that the rules won't change underneath them for the benefit of whoever's holding the pen. An AI lab publishing a protocol carries an obvious asterisk: it's their protocol, and if their competitive interests ever diverge from the spec's neutrality, they control both.

That asterisk is exactly what kills adoption at scale. Enterprise engineering leadership doesn't build production integrations on top of a single vendor's roadmap decisions dressed up as an open standard. They've been burned by that pattern before — proprietary "open" formats that quietly diverged once the sponsoring company needed differentiation. So even a genuinely well-designed protocol hits a ceiling if the governance model doesn't match the neutrality claim.

What actually changed in December 2025

The donation to the Agentic AI Foundation wasn't a rebrand. It moved spec authorship, governance, and roadmap decisions to a community body under Linux Foundation stewardship — the same governance pattern that's underpinned Kubernetes, Node.js, and dozens of other infrastructure-critical projects that needed to outlive any single sponsoring company. The moment MCP stopped being "Anthropic's protocol" is the moment it became safe to build a business on top of.

That distinction matters more than most people give it credit for. OpenAI, Google DeepMind, and Microsoft didn't need to like Anthropic to adopt MCP — they needed to trust that adopting it wouldn't hand Anthropic leverage over their own roadmaps. Vendor-neutral governance is what let a genuine competitor pick up a rival's protocol without it reading as capitulation.

The numbers only make sense in that light

By May 24, 2026, the official MCP Registry held 9,652 latest server records — knocking on the door of 10,000, five months after the foundation handoff. The Python and TypeScript SDKs combined pull roughly 97 million downloads a month. Those aren't numbers you get from a protocol that still smells like one company's internal tooling. They're numbers you get when competing labs feel safe enough to build their own tooling on the same rails, because the rails don't belong to a rival anymore.

Compare that trajectory to what tends to happen with vendor-controlled "open" specs: adoption plateaus right at the edge of the sponsoring company's own ecosystem, because nobody outside that ecosystem wants to hand over roadmap control by proxy. MCP didn't plateau there. It kept climbing right through the boundary where a closed-governance protocol usually stalls.

Why this is a story about incentives, not code

It's tempting to credit MCP's traction to good technical design — the tool/resource/prompt model is clean, the transport layer is sensible, the SDKs are genuinely pleasant to use. All true. But protocol history is littered with technically sound specs that never left their sponsor's orbit because nobody trusted the governance behind them. Good design gets you a spec people could adopt. Neutral governance is what gets people to actually bet their roadmap on it.

This is worth sitting with if you're building anything meant to become infrastructure rather than a product. The instinct to hold onto control is strong, especially when the thing you built is working. But infrastructure that stays owned by its creator has a ceiling — it can only ever be as trusted as that creator is neutral, and no single company, however well-intentioned, reads as neutral to its own competitors. Anthropic gave up formal control of MCP at almost exactly the moment that control started being worth something, which is a harder trade to make than it sounds.

What the alternative timeline probably looked like

It's worth actually imagining the counterfactual, because it makes the stakes of the December 2025 decision more concrete. Picture MCP staying formally under Anthropic's control through 2026. The technical design doesn't change — the tool/resource/prompt model is the same, the SDKs are the same quality. What changes is every enterprise architecture review that has to answer the question "who governs the roadmap of the protocol we're about to build core workflows on." In that counterfactual, the honest answer is "a company that competes with some of the vendors we also depend on," and that answer kills or slows a meaningful share of the deals that instead went through cleanly.

You can see fragments of this counterfactual already, in categories of software where a genuinely useful integration format stayed vendor-controlled and never crossed into universal adoption — it became "the standard you use if you're already in that vendor's ecosystem," rather than a genuinely neutral layer. MCP could easily have ended up there. A protocol with excellent technical design, adopted enthusiastically inside Anthropic's own developer base and by companies already committed to Claude, but treated with structural caution everywhere else. Instead, by mid-2026, MCP is closing in on 10,000 registry servers and adoption from Anthropic's most direct competitors — the opposite outcome, and the governance donation is the single clearest inflection point that explains the divergence.

The part that's easy to miss in the retrospective

None of this guaranteed MCP would win. Plenty of protocols get donated to foundations and quietly die there, governance neutrality alone doesn't manufacture demand. What the donation did was remove the one objection that no amount of technical merit could answer. Once that objection was gone, the protocol had to survive on its own design — the tool-call model, the resource abstraction, the way servers declare capabilities. That it did survive, and grew past 9,650 servers in the registry with a 950-plus-server enterprise wave building on top of it, says the underlying design was sound. The governance move just made sure the design got a fair test.

That's the pattern worth remembering: the technical bet and the governance bet are separate bets, and you can lose on either one independently. MCP is a rare case where both paid off, and the order mattered — neutral governance came first, and adoption followed. Reverse that order and you're just hoping the market forgives you for holding on too long.

Part of the "MCP One Year In" series on aiskill.market.