Linux Foundation

The Linux Foundation is a nonprofit organization that hosts open source projects and supports their governance, infrastructure and communities. Its AI work includes PyTorch, RWKV and the Agentic AI Foundation (AAIF), launched in December 2025 with contributions from Anthropic, Block and OpenAI. These projects extend its role from machine learning frameworks and models to the protocols, tools and repository guidance used by AI agents.[8][10][5]

On December 9, 2025, the Foundation launched AAIF with three inaugural projects addressing different parts of the agent stack. Anthropic’s Model Context Protocol (MCP) connects AI applications to tools and data. Block’s goose is a local-first agent framework combining language models, extensible tools and MCP integration. OpenAI’s AGENTS.md, released in August 2025, is a Markdown convention supplying coding agents with project-specific guidance across repositories and toolchains.[5][1]

AAIF was established as a directed fund under the Linux Foundation. Its governing board handles strategic investments, budget allocation, member recruitment and approval of new projects; individual projects retain autonomy over technical direction and daily operations. MCP’s existing maintainer structure and community specification-proposal process continued. Lead core maintainer David Soria Parra’s donation announcement stated that the Foundation “will not dictate the technical direction of MCP.” Anthropic also committed to continuing development investment and maintaining core infrastructure.[6]

AAIF’s launch coalition included AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI as Platinum members. At the launch, Soria Parra reported over 97 million monthly MCP SDK downloads and 10,000 active servers; the Foundation separately reported AGENTS.md adoption by more than 60,000 open source projects. These figures describe project adoption, rather than Foundation customers or unique users.[5][6]

On September 14, 2026, AAIF launched the Model Context Protocol Associate certification, its first certification, with enrollment open. MCPA is a 120-minute online, proctored, multiple-choice exam covering MCP fundamentals, architecture, execution, security and governance, and use cases. It was aligned with the July 28, 2026 specification. The announcement reported monthly downloads across MCP’s Tier 1 SDKs approaching half a billion.[15]

The Linux Foundation and PyTorch Foundation launched OpenMDW in 2025 to address AI artifacts that traditional software licenses were not designed to cover. Its framework encompasses model architecture, weights and parameters alongside code, documentation and data. On May 28, 2026, the Linux Foundation released OpenMDW-1.1 and announced NVIDIA’s plans to adopt it for future releases of Cosmos, Isaac GR00T, Ising and Nemotron. The adoption plan concerned future releases, rather than a completed migration of every existing model.[7]

These AI initiatives followed earlier work on Linux standards, developer support and shared project services. The organization’s formation brought together two nonprofits with those responsibilities.[11][12][8][5]

On January 22, 2007, the Free Standards Group (FSG) and Open Source Development Labs (OSDL) announced a merger to form the Linux Foundation, led by former FSG executive director Jim Zemlin. FSG’s Linux Standards Base provided technical specifications intended to prevent GNU/Linux fragmentation. OSDL sponsored Linus Torvalds and other kernel developers and provided shared legal support through its Linux Legal Defense Fund and Patents Commons Project. Bruce Byfield’s contemporary report described legal completion as expected in early February.[11]

Zemlin explained that the predecessors had overlapping memberships and activities and that combining them would reduce duplicated efforts. He described the Foundation’s intended role as providing “a neutral spokesperson on behalf of our members so that the platform can be represented without any one company hawking its goods.” Founding members included Hewlett-Packard, IBM, Intel, Novell, Oracle and Red Hat.[11]

The Foundation launched Linux Foundation Research in 2021 to study open source software, standards and communities. Initial work included cybersecurity, software bill of materials readiness and community collaboration. In his December 2021 update, Zemlin said non-personally identifiable research data would be available under the Community Data License Agreement—Permissive, Version 2.0. The update also credited OpenSSF’s collaboration with Harvard’s Laboratory for Innovation Science on a FOSS Contributor Survey and the Foundation’s collaboration with edX on its Jobs Report.[12]

The Foundation’s LFX platform supplied tools for operating project communities. Its 2021 offerings included EasyCLA for electronic contributor license agreements, Insights for aggregating and analyzing repository, communication-channel and contributor data, and Mentorship for recruiting and engaging participants. Project Control Center provided API-driven provisioning and configuration of resources such as cloud services and DNS.[12]

On September 12, 2022, the Foundation announced PyTorch’s move from Meta into the newly formed PyTorch Foundation. AMD, AWS, Google Cloud, Meta, Microsoft Azure and NVIDIA were founding members, and the Linux Foundation appointed Ibrahim Haddad as executive director. The transition gave the existing framework access to Foundation training, certification, research, events and LFX services. At the announcement, Mark Zuckerberg reported more than 150,000 GitHub projects built on PyTorch.[8]

LF AI & Data launched the Generative AI Commons in September 2023, naming RWKV as its first technical project entering incubation. RWKV began as an EleutherAI community project led by Bo Peng; StabilityAI sponsored compute for training the original models. The official project record credits Blink DL with contributing RWKV in September 2023. At the Commons launch, LF AI & Data reported over 55 projects, more than 60 members and over 27,000 active contributors.[10][9]

RWKV combines parallelizable training with recurrent inference without attention: the hidden state at position t is used to compute the state at t+1. In a March 2024 Latent Space episode, Eugene Cheah presented Foundation stewardship as protection against dependence on a single person, saying, “If I go rogue one day, you can just, the code will not disappear. The model will not disappear.” This was his account of the arrangement’s value, rather than a legal guarantee of perpetual availability.[9][10][4]

The Commons also proposed evaluating generative AI projects by component availability and licensing. The Model Openness Framework paper, initially submitted on March 20, 2024, was authored by Matt White, Ibrahim Haddad, Cailean Osborne, Xiao-Yang Yanglet Liu, Ahmed Abdelmonsef, Sachin Varghese and Arnaud Le Hors. Its October 2024 revision describes three ranked tiers based on completeness and openness, specifying required code, data and documentation components and their open licenses. The accompanying Model Openness Tool provides a reference implementation for evaluation.[10][13]

Before MCP joined AAIF, its coauthors discussed the trade-off between broader governance and development speed. In an April 2025 interview, they expressed concern that broader governance processes could slow development. They also described existing participation by other companies, including Microsoft’s work on the C# SDK. December’s AAIF arrangement preserved project technical autonomy and the existing MCP maintainers while establishing a board for funding and organizational decisions.[3][6]

The Foundation’s financing includes support for designated initiatives. Its December 2021 update reported that the Linux Foundation and OpenSSF had raised over $10 million for security leadership and initiatives in an October announcement. At the broader organizational level, the 2025 annual report forecast revenue above $310 million and nearly $285 million in spending supporting its mission for 2025. Those annual figures were forecasts, rather than audited closed-year results.[12][14]