There is no Google–Microsoft alliance against Anthropic and OpenAI. The thing the alliance headlines describe is Agentic Resource Discovery (ARD), an Apache-2.0 licensed v0.9 draft specification published on June 17 by eleven companies, defining how an AI agent finds a tool or a service before it calls one. It is a month old, it is a draft, and neither Google’s announcement nor Microsoft’s mentions Anthropic, OpenAI, or competition of any kind. The competitive framing belongs entirely to the coverage.
What Actually Shipped
ARD is a specification, not a consortium: posted to the Google Developers Blog and Microsoft’s Command Line blog on Wednesday, June 17, 2026, licensed Apache 2.0, developed in the open at ards-project/ard-spec, and labelled by its own repo “v0.9 (Draft). The specification is open and evolving.” No legal entity sits behind it, no membership agreement, no disclosed funding. Its repo carries 379 stars and 91 commits, modest for an industry realignment.
Eleven contributors are listed, not two. Microsoft’s own post enumerates the other ten verbatim — Cisco, Databricks, GitHub, GoDaddy, Google, Hugging Face, Nvidia, Salesforce, ServiceNow, and Snowflake — though GitHub is Microsoft-owned, so ten independent parties is the fairer count. The spec builds on the AI Catalog data model from the Linux Foundation’s AI Catalog Working Group.
Where the “Alliance” Came From
The Information ran “Google and Microsoft Team Up to Beat Back Anthropic and OpenAI” (Efrati, Holmes, Perloff) around June 18, the day after the spec appeared; it is paywalled, and I cannot rule out that it contains sourcing I have not seen. Two days later CryptoBriefing published “Google, Microsoft, and Salesforce back new AI software standard to counter OpenAI and Anthropic,” citing no external reporting at all for the word “counter.” That the frame landed twice in outlets not downstream of each other is the tell, and it predates ARD anyway — CNBC was already running “Microsoft and Google take on Anthropic and OpenAI in AI coding models” on June 1, about the two firms separately chasing the labs on coding. It is a template, and ARD was the next set of facts it got stamped onto.
The damage shows downstream. A buildfastwithai roundup titled “AI News Today July 13 2026,” published four weeks after the announcement it describes, reports that the contributors “agreed to support a shared technical standard for connecting AI agents to business software, positioning it as a counter to Anthropic’s Model Context Protocol.” No primary source says that; every primary source contradicts it. Dates drift both directions too — other coverage puts ARD in “late May 2026,” and InfoQ’s July 14 piece frames it as news without giving any announcement date at all — the same failure mode that invented a launch date for Gemini 3.5 Pro.
ARD Does Not Compete With MCP
MCP is how an agent invokes a tool; ARD is how it finds the tool beforehand. Hugging Face’s launch post says it about as plainly as a contributor can: “MCP gives agents a standard way to call tools… The Agentic Resource Discovery (ARD) specification is the discovery layer that sits in front of them.” ARD catalogs explicitly describe MCP servers, A2A agents and OpenAPI tools, and the agent then connects over the resource’s native protocol — MCP, A2A, or REST. Snowflake is blunter: “the discovery service never sits in the invocation path.” ARD sits above the tool-use layer, not across from it.
The frame also collapses on governance. Google and Microsoft are both Platinum members of the Agentic AI Foundation, the Linux Foundation fund formed December 9, 2025 that is the neutral home of Anthropic’s donated MCP — sitting alongside Anthropic and OpenAI, Platinum members too. ARD borrows that foundation’s working-group data model, so it is not routing around Linux Foundation governance either. Nor are the two firms aligned generally: TechCrunch, sourcing Bloomberg, reported on July 15 that Microsoft trains salespeople to talk down Google’s AI products in the same breath as OpenAI’s and Anthropic’s, having spent the spring shipping seven in-house MAI models of its own. That is not a bloc. It is a company going it alone and signing a spec on the way past.
The strongest competitive case still deserves stating: a discovery layer that ranks which tools an agent reaches sits upstream of the layer that invokes them, and whoever owns ranking owns a great deal without ever touching the call. That is a defensible argument about effects rather than design — and not what any of the headlines said.
The Part That Is Actually Useful
Strip the framing off and the mechanism is familiar to anyone who has edited a robots.txt. You publish a machine-readable ai-catalog.json at a well-known path on your own domain, listing what your business exposes — APIs, services, MCP servers, booking endpoints. Federated registries crawl and index those catalogs, agents send natural-language discovery queries and get ranked matches back, then connect straight to whatever they picked. There is no central catalog and no gatekeeper to approve you, because domain ownership is the trust anchor: your website becomes the verification root for your own capabilities, the way it already is for your sitemap. For a business that publishes capabilities rather than merely buying AI tools, that is a plausible successor to the llms.txt and GEO play, with a better trust model than either. The reported reference implementations are GitHub’s Agent Finder in Copilot, Hugging Face’s Discover Tool, and Google’s Agent Registry inside the Gemini Enterprise Agent Platform — reported by InfoQ and developer-tech, not by anything I have tested.
Why Not To Build It Yet
It is a v0.9 draft one month old whose contributors are still taking changes through pull requests: correct for a draft, poor as a foundation for your Q3 roadmap. More decisively, no AI client is listed anywhere as having native ARD support. The spec site’s connection guides name Claude, ChatGPT, Copilot and Gemini, and Snowflake’s launch post promises that “ARD-compatible AI interfaces, including Snowflake CoWork, Claude, Copilot or your own application, can find and invoke the same Cortex Agent without re-registration” — but that is a contributor’s forward-looking marketing copy, not evidence that anything parses ai-catalog.json today. Publishing a catalog now bets on a client population nobody has shown exists.
The precedent is poor. Ahrefs studied all 137,210 domains in its Web Analytics with traffic in May 2026, found roughly 38,000 carrying a valid llms.txt, and of those ~38,000, 97% received zero llms.txt requests that month — and Ahrefs flags its own adoption figure as an upper bound, since its customers skew SEO-aware. Google’s John Mueller has been publicly cool on llms.txt for related reasons: it cannot help a model decide which site to visit, though he allows that “if someone is already on your website, maybe some kind of automated system is helpful.” Mueller was talking about llms.txt, not ARD, and nobody should put ARD words in his mouth. What transfers is the shape of the risk — a file nothing reads is not a strategy.
The One Competitive Read That Holds
Anthropic and OpenAI are not on the contributor list, neither has announced client-side support, and ARD sits outside the Agentic AI Foundation track even while borrowing that foundation’s data model. There is no evidence either lab was invited and declined. But the absence is real, and it points at a real structural tension: the platform and data companies want the layer where agents find things, the labs own the layer where agents do things and the assistant surface where the user sits, and those are different businesses with a legitimate reason to pull apart. That is an interpretation — mine, and worth what an interpretation is worth. What is not an interpretation is that no primary source shows an alliance, that ARD is Apache-2.0 with open governance and no legal entity, that both firms share a foundation board with the labs they supposedly oppose, and that Microsoft is selling against Google this quarter. None of that proves a pact does not exist; it is a large pile of evidence against one, which is the most any negative claim gets.
So understand ARD, watch whether a client you actually use ever ships native support, and do not publish a catalog for an audience of zero. Check the contributor list, the license, the version number and the governing body before believing the next standards war — the same habit that catches a lab’s own chart showing its new model losing. It took about ten minutes here, and the story did not survive it.
Key Details
| Spec | Detail |
|---|---|
| Specification | Agentic Resource Discovery (ARD) |
| Published | June 17, 2026 (Google Developers Blog, Microsoft Command Line) |
| Status | v0.9 — Draft |
| License | Apache 2.0 |
| Contributors | 11: Cisco, Databricks, GitHub, GoDaddy, Google, Hugging Face, Microsoft, Nvidia, Salesforce, ServiceNow, Snowflake |
| Built on | AI Catalog data model (AI Catalog Working Group, Linux Foundation) |
| Mechanism | ai-catalog.json at a well-known path; federated registries crawl and index it |
| Trust anchor | Domain ownership — no central catalog, no gatekeeper |
| Relation to MCP | Complementary: ARD is discovery, MCP is invocation |
| Anthropic / OpenAI | Not contributors; no announced client-side support |
| Repo activity | 379 stars / 46 forks / 91 commits / 28 open issues / 6 PRs (as of 2026-07-16) |
| Repository | ards-project/ard-spec |
Sources
- Announcing the Agentic Resource Discovery specification — Google Developers Blog
- Agentic Resource Discovery specification (ARD) — Microsoft Command Line
- ARD specification repository — GitHub
- Agentic Resource Discovery — official site and contributor list
- The Agentic Resource Discovery launch — Hugging Face
- Agentic Resource Discovery specification — Snowflake
- Linux Foundation announces the formation of the Agentic AI Foundation
- Google, Microsoft back draft AI agent discovery spec — Search Engine Journal
- Agentic Resource Discovery spec — InfoQ
- GitHub and Google back ARD AI agent discovery — Developer Tech
- Google and Microsoft Team Up to Beat Back Anthropic and OpenAI — The Information (paywalled)
- Google, Microsoft, and Salesforce back new AI software standard to counter OpenAI and Anthropic — CryptoBriefing
- Microsoft and Google take on Anthropic and OpenAI in AI coding models — CNBC
- Microsoft is reportedly training salespeople to talk down OpenAI and Anthropic — TechCrunch
- Microsoft replaces OpenAI, Anthropic with own AI in some apps — Bloomberg
- AI News Today July 13 2026 — buildfastwithai
- The llms.txt study — Ahrefs
- Google’s Mueller says llms.txt can’t help LLMs differentiate sites — Search Engine Journal