Google, Microsoft back ARD spec to counter Anthropic's MCP

The standards fight over enterprise AI agents just picked up a new front. On June 17, 2026, eleven companies published a draft specification called Agentic Resource Discovery, or ARD, aimed at how agents find the tools, APIs, and MCP servers they need. The Information framed it plainly: Google and Microsoft are teaming up to push back on Anthropic and OpenAI in the plumbing layer that is quietly deciding who owns enterprise AI workflows.

The coalition matters as much as the spec. Google and Microsoft anchor it, joined by Salesforce, Snowflake, ServiceNow, Cisco, Databricks, GitHub, GoDaddy, Hugging Face and Nvidia. The mechanism is drier than the politics. Each organization publishes an ai-catalog.json manifest at a well-known path on its own domain, listing the tools, MCP servers, agents or APIs it exposes, and federated registries crawl those catalogs to answer plain-language discovery queries from agents. The spec is Apache 2.0 licensed and builds on an AI Catalog data model maintained by a Linux Foundation working group.

The absence is the tell. Anthropic, whose Model Context Protocol has become the de facto tool-connection standard, is not a backer. Neither is OpenAI. ARD does not replace MCP; it sits above it as a discovery layer, and the published guides treat Claude, ChatGPT, GitHub Copilot, Microsoft Copilot and Gemini as ARD clients rather than as ARD authors. In practice, that lets the incumbents decide where agents look first, before any model vendor's protocol gets invoked.

The honest caveat is how early this is. ARD is a v0.9 draft, and one tracker noted that as of mid-June, none of a sampled 39 top sites were yet publishing an ai-catalog.json. What the reporting does not give you is a firm shipping timeline from each backer, whether Anthropic or OpenAI will counter with their own discovery layer or eventually fold in under enterprise pressure, or how ARD's governance will interlock with A2A, which also lives at the Linux Foundation.

The forward-looking read is straightforward. If ARD works, the value accrues to whoever runs the registries and hosts the catalogs, which is exactly the ground the cloud and SaaS incumbents already own. That is the point.