12 Dec 2025 AI Engineering5 min read

The Protocol Layer Goes Neutral

Through the Agentic AI Foundation, the Linux Foundation is moving MCP, goose, and AGENTS.md into neutral governance. We examine which risk the transfer removes, which technical gaps remain, and what engineering teams should do in response.

A Standard Owned by One Vendor

Anthropic released the Model Context Protocol on November 25, 2024, as an open standard connecting AI models to tools and data. The JSON-RPC-based client-server protocol lets a model-side client discover and use tools, resources, and prompts exposed by servers. Twelve months later, the numbers establish its reach: more than 10,000 published MCP servers, over 7 million monthly downloads of the official Python and TypeScript SDKs, and support in Claude, ChatGPT, Gemini, Microsoft Copilot, Cursor, and VS Code.

That adoption made MCP the de facto standard, but not yet a neutral one. Until this week, Anthropic owned the specification, repositories, and trademark. Contributions from Google, Microsoft, or Block therefore strengthened a competitor's property, while roadmap decisions ultimately remained with one company. Only goodwill prevented a future change of license or direction. That risk may be acceptable for a library. For infrastructure on which an industry standardizes, it is a structural flaw.

An agent reaches its tools over MCP — one standard surface. 1/4

What Happened on December 9

On December 9, 2025, the Linux Foundation announced the formation of the Agentic AI Foundation (AAIF), co-founded by Anthropic, Block and OpenAI with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. Eight Platinum members anchor the fund; Gold and Silver members range from Cisco and IBM to Hugging Face and Zapier. Three projects were donated at launch: MCP, goose and AGENTS.md.

The operative word is donation, not partnership: ownership transfers to the foundation. A directed fund gives an industry consortium its own budget, membership, and technical governance under the Linux Foundation's legal umbrella. The same organization stewards the Linux kernel, Kubernetes, Node.js, and PyTorch. MCP's operating process is nevertheless intended to remain stable; Anthropic says the maintainers and community-driven decision process will continue unchanged.

Three Donations Define the Stack

The three founding projects are not arbitrary. They cover three distinct layers of an agentic system: MCP standardizes how agents reach tools and data, goose is a working open-source agent that consumes that protocol, and AGENTS.md standardizes how repositories instruct coding agents. Together they form a minimal, fully open reference stack.

None of the three is a model. That is the point. The AAIF governs the connective tissue — protocol, runtime, convention — while the models above it remain proprietary and competitive. The foundation draws a line through the stack: below it, cooperation; above it, competition.

Project	Donor	First released	Layer
MCP	Anthropic	November 2024	Protocol: connects models to tools and data; 10,000+ published servers
goose	Block	Early 2025	Runtime: local-first open-source agent framework built on MCP
AGENTS.md	OpenAI	August 2025	Convention: per-repository agent instructions; 60,000+ adopting projects

Why Neutrality Is an Engineering Concern

Neutral governance changes three concrete conditions. First, the foundation holds the intellectual property — specification, repositories, and trademarks — so no single member can relicense it. Second, changes follow documented governance rather than one vendor's roadmap. Third, incentives shift: OpenAI engineers can improve MCP without increasing the value of an Anthropic asset, and vice versa.

The pattern has precedent. Google moved Kubernetes into the newly created CNCF in 2015; the project became an industry default partly because no competitor had to build on Google's property. The Node.js Foundation, also 2015, ended the io.js fork. Foundation transfer is the normal end state of successful open infrastructure. MCP reached it in just over a year.

For enterprise buyers, neutrality is a procurement criterion, not a sentiment. A standard held by a foundation survives the strategy changes of any single company. That is the difference between adopting a protocol and adopting a product.

What the Foundation Does Not Solve

The transfer changes ownership, not technology. MCP's open problems therefore remain: prompt injection through tool descriptions and server responses is unsolved, quality varies widely across the 10,000-plus published servers, and no certification separates sound implementations from abandoned projects. The 2025 specification revisions improved authorization, but securing a deployment remains the operator's responsibility.

AGENTS.md standardizes where instructions live, not how agents interpret them; two agents reading the same file can still behave differently. Governance also has a price: foundation consensus is slower than single-vendor decisions, and the Linux Foundation hosts stalled projects alongside thriving ones. Neutrality buys stability. It does not buy velocity.

Fragmentation above the protocol also persists. Orchestration frameworks remain incompatible; the AAIF standardizes how agents reach tools, not how teams build, evaluate or supervise agents. Anyone expecting a full-stack standard will be disappointed. The foundation governs plumbing, deliberately.

Consequences for Engineering Teams

For adopters, the main change is a simpler risk calculation. An integration implemented as an MCP server serves every compliant client, while the protocol beneath it can no longer be relicensed unilaterally. The practical response is clear: expose internal systems as MCP servers rather than platform-specific plugins, maintain an AGENTS.md in each repository, and separate orchestration logic cleanly from protocol bindings. The layers above the protocol remain proprietary and volatile.

At Blue IT Systems, we standardized client integrations on MCP during 2025. The residual risk was governance rather than technology: a protocol owned by one vendor can change direction beneath an existing architecture. Since December 9, that risk is materially smaller. We no longer see a reason to maintain parallel integration paths solely as a hedge against protocol ownership.

Outlook

The next stage is already visible. The November 25, 2025 specification release — asynchronous operations, statelessness, server identity, and official extensions — will propagate through the SDKs in early 2026. After that, the young standard chiefly needs testable infrastructure: a conformance suite, enterprise security profiles, a more mature official server registry, and further project donations under the shared umbrella.

We draw two predictions from this direction. First, by the end of 2026 MCP support will be a checkbox rather than a differentiator; competition will move up the stack into orchestration, evaluation, and security. Second, "MCP-compatible" will become a verifiable claim only once a conformance program exists. Pace remains the open question: a foundation improves a standard's chances of survival, but does not guarantee technical leadership. The next twelve months will show whether neutral governance can keep up with agentic AI.