The Agentic AI Foundation (AAIF) is a directed fund under the Linux Foundation that provides vendor-neutral governance for open standards and open-source software used by AI agents. It was announced on December 9, 2025 by Anthropic, Block, and OpenAI, with platinum support from Google, Microsoft, AWS, Cloudflare, and Bloomberg. The foundation launched with three flagship projects: the Model Context Protocol (MCP), donated by Anthropic; goose, donated by Block; and AGENTS.md, donated by OpenAI.[1][2][3]
The foundation operates from aaif.io and describes itself as a neutral home built on transparency, collaboration, and standardization to advance the public interest in agentic AI. Linux Foundation executive director Jim Zemlin said bringing the projects together "ensures they grow with the transparency and stability that only open governance provides."[2][3]
Agentic AI tooling expanded quickly through 2024 and 2025. MCP, originally announced by Anthropic in November 2024 as an open standard for connecting AI models to tools and data sources, reached more than 10,000 active public servers and roughly 97 million monthly SDK downloads (Python and TypeScript combined) in its first year, with first-class client support from ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code.[1][7] The Claude connector directory alone listed more than 75 servers at the time of the donation.[1]
That growth made the question of who governs the protocol a real one. Anthropic had already moved MCP from a single-vendor specification to a multi-stakeholder governance model earlier in 2025, but the project was still hosted by Anthropic. As GitHub's Martin Woodward put it, once the protocol started showing up in production at multiple vendors, "the need for neutral governance became unavoidable."[8] Linux Foundation hosting addresses three concerns at once: legal neutrality for the trademark and codebase, predictable governance for downstream adopters, and a venue where competitors can contribute without giving any one of them control.
The AAIF launch followed a similar move six months earlier. In June 2025, Google donated the Agent2Agent (A2A) protocol to the Linux Foundation, creating a separate Agent2Agent project for inter-agent communication.[10] MCP and A2A are complementary rather than competing: MCP connects agents to tools and data, while A2A coordinates agents talking to other agents.
The foundation launched with eight platinum members and a much larger tier of gold and silver supporters.
| Tier | Member |
|---|---|
| Platinum | Amazon Web Services |
| Platinum | Anthropic |
| Platinum | Block |
| Platinum | Bloomberg |
| Platinum | Cloudflare |
| Platinum | |
| Platinum | Microsoft |
| Platinum | OpenAI |
| Gold (selected) | Adyen, Arcade.dev, Cisco, Datadog, Docker, Ericsson, IBM, JetBrains, Okta, Oracle, Salesforce, SAP, Shopify, Snowflake, Temporal, Twilio |
| Silver (selected) | Apify, Hugging Face, Mirantis, Pydantic, SUSE, Uber, WorkOS, Zapier |
Three of the eight platinum members (Anthropic, Block, and OpenAI) also contributed code or specifications as founding projects. The other five entered as financial backers and adopters.[2][3]
The AAIF launched with a deliberately small project portfolio: one wire protocol, one agent runtime, and one repository convention. Together they cover the layers an autonomous coding or research agent typically needs.
| Project | Donor | Released | What it does |
|---|---|---|---|
| Model Context Protocol (MCP) | Anthropic | November 2024 | Open standard for connecting LLM applications to external tools, data sources, and services through a uniform interface. |
| goose | Block | Early 2025 | Open-source, local-first agent framework that runs on a developer's machine, with a desktop app, CLI, and API; supports 15+ model providers. |
| AGENTS.md | OpenAI | August 2025 | Plain Markdown convention for a project-level instruction file that AI coding agents read for build steps, tests, and conventions. |
MCP is the largest of the three by downstream adoption. It defines how an AI client discovers and calls tools exposed by a server, how it reads resources, and how it exchanges prompts. By the time of the donation, MCP servers existed for issue trackers, code search, observability platforms, databases, and most major cloud APIs. The protocol's governance model, introduced earlier in 2025, did not change with the move to AAIF: the existing maintainers continue to run technical decisions, with the Linux Foundation providing legal and infrastructure backing rather than steering the spec.[1][6]
Anthropic's chief product officer Mike Krieger said the goal of donating MCP was to keep it "open, neutral, and community-driven as critical infrastructure."[3] AWS vice president Swami Sivasubramanian framed it from the adopter's perspective: a vendor-neutral foundation gives developers confidence to invest in MCP as a universal standard rather than worry about future licensing or direction changes.[3]
goose is a general-purpose, local-first agent framework that Block (the parent of Square, Cash App, Afterpay, and Tidal) released in early 2025. It runs entirely on a user's own machine, offers a native desktop app for macOS, Linux, and Windows, plus a CLI and an embeddable API, and works with more than 15 model providers including Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure, and Bedrock.[5][9] goose is also one of the most extensive consumers of MCP, which made it a natural reference implementation for the protocol.
The project's home moved from github.com/block/goose to the AAIF organization on GitHub in April 2026, four months after the foundation announcement.[9]
AGENTS.md is the lightest of the three contributions. It is just a CommonMark file convention: a project drops an AGENTS.md at the repository root (or in subdirectories), and AI coding tools that respect the format read it for build instructions, test commands, and house style. There is no schema, no YAML front matter, and no required vocabulary; the file is plain Markdown with conventional headings such as ## Build & Test and ## Code Style that agents scan as semantic cues.
OpenAI introduced AGENTS.md in August 2025, originally to give Codex a predictable way to find project-specific instructions. By the time it joined AAIF in December 2025, more than 60,000 open-source projects had adopted the convention, including support from Amp, Codex, Cursor, Devin, Factory, Gemini CLI, GitHub Copilot, Jules, and VS Code.[3][4]
AAIF uses the Linux Foundation directed-fund model, which separates business governance from technical governance.
| Body | Composition | Responsibilities |
|---|---|---|
| Governing Board | Representatives from platinum members, plus elected seats for gold members | Strategic direction, budget, member recruitment, approval of new projects, marketing, events |
| Technical Committee | Project maintainers and elected technical representatives | Cross-project technical coordination, project incubation criteria, technical standards |
| Per-project maintainer groups | Each project keeps its existing governance | Day-to-day technical decisions, releases, RFCs, code review |
Individual projects retain full autonomy over their technical direction. The MCP maintainers, for example, continue to make protocol decisions; AAIF provides a neutral legal home, infrastructure, and a forum for cross-project work, not a top-down technical authority.[6]
In February 2026, the foundation announced David Nalley, AWS director of developer experience, as governing board chair. Nalley is also a long-time Apache Software Foundation board member, which gave the appointment some additional credibility with the open-source community.[4]
AAIF is part of a broader pattern of the Linux Foundation acting as a neutral host for industry-critical infrastructure. Three earlier sub-foundations make useful reference points.
| Foundation | Founded | Anchor projects | Domain |
|---|---|---|---|
| AAIF | December 2025 | MCP, goose, AGENTS.md | Open standards and software for AI agents |
| CNCF (Cloud Native Computing Foundation) | 2015 | Kubernetes, Prometheus, Envoy, etcd | Cloud-native and container infrastructure |
| OpenJS Foundation | 2019 | Node.js, jQuery, Electron, Webpack | JavaScript runtimes, libraries, and tooling |
| Agent2Agent Project | June 2025 | A2A protocol | Inter-agent communication (donated by Google) |
The CNCF analogy comes up most often in coverage of AAIF. CNCF rose alongside Kubernetes and standardized container orchestration after a brief period of vendor-controlled alternatives such as Docker Swarm and Mesos. Observers, including GitHub's open-source maintainers blog, have framed AAIF as following the same arc for agentic AI: a moment where one or two protocols are clearly winning, and the industry collectively decides to put them in a neutral home before any single vendor consolidates control.[8]
OpenJS provides a slightly different precedent. It formed in 2019 by merging the Node.js Foundation and the JS Foundation, primarily to cut down on overlap and streamline JavaScript ecosystem governance. AAIF started as a single new entity rather than a merger, but its multi-project structure (one wire protocol, one runtime, one specification) borrows the same idea: a small umbrella with technically autonomous projects underneath.
MCP adoption was already large at the time of the AAIF announcement and has continued to grow. The headline figures Anthropic shared on December 9, 2025 included roughly 97 million monthly SDK downloads, more than 10,000 active public servers, and adoption by every major commercial AI assistant.[1] GitHub's 2025 Octoverse report independently noted 1.13 million public repositories importing LLM SDKs (a 178% year-over-year jump) and over a million pull requests authored by Copilot agents alone, both of which depend on protocols like MCP for tool integration.[8]
AGENTS.md crossed 60,000 projects before joining the foundation and continued to spread through coding-agent vendors during the first quarter of 2026.[3] goose's adoption is harder to measure precisely, but Block reported "thousands of developers" using it by the time of the donation.[9]
On February 24, 2026, the Linux Foundation announced 97 new AAIF members: 18 new gold members and 79 new silver members. The list included Akamai, American Express, Autodesk, Circle, Diagrid, Equinix, Global Payments, Hitachi, Huawei, Infobip, JPMorgan Chase, Keycard, Lenovo, Red Hat, ServiceNow, TELUS, UiPath, and Workato at the gold tier, and Mistral AI, Neo4j, Airbyte, and 1Password among the silver tier.[4]
Total membership reached 146 organizations roughly two and a half months after launch. The same announcement confirmed David Nalley's appointment as governing board chair.[4]
The foundation announced two MCP Dev Summits within its first quarter of operation:
| Event | Dates | Location |
|---|---|---|
| MCP Dev Summit North America | April 2-3, 2026 | New York City |
| MCP Dev Summit Bengaluru | June 9-10, 2026 | Bengaluru, India |
Both summits focus on protocol roadmap discussions, security and authentication patterns, and ecosystem development.[4]
The practical alternative to MCP-plus-AGENTS.md-plus-goose is one of the proprietary agent stacks shipped by major model vendors. Each handles tool calling, planning, and orchestration, but with vendor-controlled formats and pricing.
| Stack | Vendor | Tool integration | Lock-in profile |
|---|---|---|---|
| AAIF projects (MCP + goose + AGENTS.md) | Linux Foundation | MCP servers, vendor-neutral | Open spec, BSD/Apache licensed code |
| Assistants API and Responses API | OpenAI | Native function calling, file search, code interpreter | Tied to OpenAI hosting and pricing |
| Claude Agent SDK | Anthropic | Native tool use, MCP for external tools | Tied to Anthropic API; uses MCP for portable parts |
| Vertex AI Agent Builder | Native function calling, A2A for agent-to-agent | Tied to Google Cloud | |
| AWS Bedrock Agents | Amazon | Action groups, knowledge bases | Tied to Bedrock |
MCP changes this picture for the tool-integration layer specifically. A connector written once as an MCP server can be consumed by Claude, ChatGPT, Cursor, Gemini, Copilot, and any other MCP-compatible client. That removes one of the largest sources of platform-specific work in agent development. The other layers (model choice, planning, billing, security) remain vendor-decisions.
AAIF is not without skeptics, and the criticism splits into a few recognizable arguments.
The first is project maturity. Hacker News commenters and InfoQ's coverage both flagged that MCP was barely a year old at the time of donation, which is unusually early for a Linux Foundation transition. CNCF's anchor project Kubernetes had two years of Google production use and a substantial external contributor base before donation; MCP went from initial Anthropic spec to LF directed fund in roughly thirteen months.[6]
A second concern is corporate concentration. All eight platinum founders are large US-headquartered tech firms or financial institutions with significant overlap in commercial interests. Critics argue this looks less like a neutral standards body and more like a coordination mechanism for incumbents. Foundation supporters counter that the alternative was a single-vendor protocol with the same sponsors as customers; a multi-vendor foundation at least distributes formal control.
The third tension is standards versus competition. OpenAI, Anthropic, Google, and AWS all sell competing agent platforms. Putting shared infrastructure in a neutral foundation reduces some friction, but it does not resolve the underlying tension: the same companies that contribute to MCP also have proprietary agent products that benefit from non-portable features. How aggressively AAIF projects can evolve when major contributors disagree commercially is still untested.
Finally, there is the question of revenue sustainability. Linux Foundation sub-foundations typically fund themselves through membership dues, training, certification, and events. InfoQ's coverage asked whether MCP and AGENTS.md, which are specifications more than software products, can sustain the events-and-certification model that worked for Kubernetes and Linux itself.[6] The early MCP Dev Summits and the rapid membership growth suggest the model is at least viable in the short term.
None of these critiques are fatal, but they are worth keeping in mind. Open governance is a process, not an outcome, and AAIF will be judged over years rather than the few months it has so far existed.