On May 21, 2025, OpenAI announced native support for remote MCP servers in its Responses API and joined MCP’s steering committee—while MCP is a protocol open-sourced in November 2024 by Anthropic, OpenAI’s most direct competitor. Prior to this, OpenAI had already added MCP support to its own agent development toolkit. On December 9 of the same year, OpenAI, Anthropic, and Block jointly announced the formation of the Agentic AI Foundation under the Linux Foundation—and Anthropic donated the entire MCP protocol to it.
It is already unusual for a competitor to adopt your standard; it is even more unusual for that competitor to co-host the standard with you. What exactly happened over these thirteen months is worth every AI tool user’s attention—because today, nearly every external tool you integrate into your AI assistant likely travels via this protocol.
## The Problem It Solves: The N×M Integration Quagmire
MCP stands for Model Context Protocol. According to Anthropic’s November 2024 announcement, it tackles a longstanding issue: no matter how intelligent a model is, it remains trapped behind information silos—each new data source requires custom integration code. The protocol’s approach compresses the pairwise adaptation of “N applications to M tools” into “everyone speaking the same interface”: tool providers write one MCP server, application developers write one MCP client, and no more point-to-point glue code is needed in between. David Soria Parra, one of MCP’s co-creators, later put this goal more bluntly to TechCrunch: developers build once, and their tool works across all clients.
Architecturally, per the MCP specification, it follows a three-layer “client–host–server” model: the host (e.g., a chat application) creates and manages multiple clients; each client maintains a stateful, one-to-one session with a server; messages are exchanged via JSON-RPC. During initialization, both client and server declare the features they support, and those declared capabilities define what protocol features may be used within that session. The specification lists four design principles—the third, most relevant to ordinary users, states: servers cannot read the full conversation history nor observe other servers’ activities—the dialogue history remains solely with the host, and security boundaries are enforced by the host. In other words, each tool you integrate, in principle, cannot access the full conversation or see what other tools are doing.
## A Thirteen-Month Timeline
This timeline draws exclusively from official announcements and documentation issued by the involved parties. In November 2024, Anthropic open-sourced the protocol specification, SDKs, and a set of prebuilt servers (for Google Drive, Slack, GitHub, Postgres, etc.); Block and Apollo integrated MCP into their systems, and developer-tool companies including Zed, Replit, Codeium, and Sourcegraph also began adoption. On May 21, 2025, OpenAI announced native support for remote MCP servers in its Responses API—stating, in its own words, that developers can connect models to any MCP-server-hosted tool using just a few lines of code. Its announcement listed popular servers including Cloudflare, HubSpot, PayPal, Shopify, Stripe, and Zapier, and explicitly stated that invoking such tools incurs no additional charge. According to a Linux Foundation press release dated December 9, 2025, MCP had by then been adopted by platforms including Claude, Cursor, Microsoft Copilot, Gemini, VS Code, and ChatGPT; over 10,000 public MCP servers were available, spanning developer tools all the way to Fortune 500 enterprise deployments.
The protocol itself has evolved rapidly. Its version numbers are date strings, incremented only upon breaking changes: after the initial release on 2024-11-05, versions followed on 2025-03-26, 2025-06-18, 2025-11-25, and the current version, 2026-07-28. According to Anthropic’s donation announcement, the November 25, 2025 version introduced asynchronous operations, stateless mode, server identity, and an official extension mechanism; MCP also launched an official community server registry. The current version embeds version declarations in every request, adds a mandatory `server/discover` endpoint enabling clients to query a server’s supported versions, capabilities, and identity in a single call, and sets a minimum retention period of twelve months (ninety days for fast-track deprecations) for deprecated features. In scale, Anthropic disclosed that monthly downloads of its Python and TypeScript SDKs exceed 97 million.
## What Changed After Donating to the Foundation
The December 9, 2025 donation marked a turning point in this timeline. Anthropic’s Chief Product Officer Mike Krieger stated at the time of donation: when MCP was open-sourced, the team merely hoped other developers would find it useful; one year later, it had become the industry standard for connecting AI systems with data tools. The Agentic AI Foundation is a directed fund under the Linux Foundation, with three founding projects: Anthropic’s MCP, Block’s goose, and OpenAI’s AGENTS.md. As reported by TechCrunch, the Foundation specifically targets standardization and orchestration for the agent era—including shared security patterns and interoperability. OpenAI likewise acknowledged in its own announcement that it had been an early adopter and core contributor to MCP, and had built it into the foundation of ChatGPT’s connectors and applications. Eight platinum members have joined: AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI—essentially the entire roster of established players in the AI infrastructure space.
What merits close scrutiny are governance details. As explained by core maintainer Soria Parra on the official MCP blog, “very little changed” post-donation: the original maintainers still steer the protocol’s direction; community input continues to follow the SEP (Standard Enhancement Proposal) process; the Linux Foundation provides a neutral home and infrastructure but does not dictate technical direction; the Foundation’s board oversees strategic investment, budgeting, membership, and approval of new projects—but each project retains full autonomy over its technical roadmap. Anthropic’s donation announcement puts it even more directly: MCP’s governance model remains unchanged, and maintainers will continue prioritizing community input and transparent decision-making. This “neutral home” has precedent: Kubernetes, PyTorch, and Node.js all reside under the Linux Foundation. Linux Foundation CEO Jim Zemlin told TechCrunch: funding does not equal control—the roadmap belongs to the Technical Steering Committee, and no single member can unilaterally decide.
## My Three Judgments
Below are my synthesized judgments drawn from consolidating these materials—not verbatim quotes from any single announcement.
First, competitors rushing to adopt MCP is not industry camaraderie, but a distribution strategy. MCP’s first-mover advantage lies in its server ecosystem: ten thousand servers mean any latecomer building its own protocol must start from zero in persuading tool providers. For OpenAI and others, the cost of adoption is “relinquishing control over interface definition”; the cost of non-adoption is “individually adapting every tool in the ecosystem—adapting to someone else’s standard.” The latter is far more expensive, so the choice is, in fact, straightforward.
Second, the donation changes exit cost—not day-to-day control. Governance documents make this clear: maintainers remain unchanged, processes remain unchanged. What the Foundation truly guarantees is “this protocol won’t die in the hands of a single company”—a critical requirement for enterprises betting production systems on the protocol. Yet judging the authenticity of neutrality still hinges on the board’s composition: all eight platinum members are entrenched stakeholders in the AI ecosystem—neutrality resides in procedure, not in interest.
Third, the specification places the trust boundary at the host—a smart move, and also an act of responsibility-shifting. “Servers cannot see the full conversation and are isolated from one another” means the protocol itself does not defend against attacks where “data becomes command”; it delegates that defense to the host application and the user authorization layer. The protocol handles wiring—not human intent. Whether an integrated tool is trustworthy ultimately remains the user’s own judgment.
## Key Takeaways
If you’re simply using an AI assistant: before integrating third-party tools, review the host’s authorization prompts—because the protocol places responsibility for security boundaries squarely on the host and on you. If you’re developing: MCP’s date-based versioning and backward-compatibility commitments mean upgrade costs are manageable—but the “twelve-month” deprecation policy is a floor, not a guarantee. If you care about industry dynamics: top companies have entrusted a critical interface to a neutral foundation—will it become a commons like Kubernetes, or evolve into another kind of club good? The board’s decisions over the next two years will speak louder than its founding declaration.
Which MCP tools have you integrated—and what glitches have you encountered? Feel free to share in the comments.
Source: Linux Foundation Agentic AI Foundation Launch Press Release https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation