The agent standard Google and Microsoft just backed runs on top of Anthropic's MCP

The Information reported over the weekend that Google, Microsoft, Salesforce, Snowflake and ServiceNow have agreed to support a shared protocol for connecting AI agents to enterprise software, and the story was framed as the incumbents ganging up on Anthropic and OpenAI. That framing is doing a lot of work. The specification these companies actually signed is public, it has been public since June 17, and it does not replace Anthropic's Model Context Protocol. It sits above MCP and calls into it.

The spec is Agentic Resource Discovery, ARD, and Snowflake's own announcement describes it plainly: an open protocol for how agents and tools "are cataloged, searched and discovered across an enterprise," developed with Microsoft, GoDaddy and others. The published supporter list around the June launch also included Cisco, Databricks, GitHub, Hugging Face and NVIDIA. It is Apache 2.0.

What ARD actually does

Read the docs and the mechanism is unglamorous, which is usually a good sign.

A company publishes a static JSON file called ai-catalog.json at https://yourdomain.com/.well-known/ai-catalog.json. Each entry names a resource (an MCP server, an API, an agent), gives it a domain-anchored identifier in the form urn:ai:acme.com:server:weather, states its type and URL, and lists two to five plain-English example queries so a semantic search index has something to embed. If you cannot host at .well-known, you point a DNS TXT record at the file instead.

From there the flow has four steps. A publisher describes its resources in the manifest. A discovery service curates a collection by crawling those manifests or ingesting an internal inventory. A client sends the discovery service a natural-language query and gets back ranked entries with schemas and endpoints. Then the client connects directly to the chosen resource over that resource's own protocol, which Snowflake lists as MCP, A2A or REST.

That last step is the one worth pausing on. The discovery service never sits in the invocation path. It hands you an address and gets out of the way. Authentication and data access stay between the client and the agent. So an enterprise that adopts ARD in full is still making its actual tool calls over MCP.

The fight is over the registry, not the wire

Calling ARD a rival to MCP is like calling DNS a rival to TCP. They are answers to different questions. MCP is how an agent talks to a tool once it knows the tool exists. ARD is how it finds out the tool exists at all.

Which does not mean the discovery layer is a small prize. A registry decides what gets indexed and how results are ranked, and Snowflake's post says the quiet part directly: "because AI tools only surface what the registry indexes, the registry becomes the place where governance capabilities and approval decisions are reflected." Whoever runs the endpoint your employees' agents query controls which agents they can see. If the default discovery service in Microsoft 365 or Google Workspace ranks first-party tools above everything else, no one has to block a competitor to beat it.

The real problem ARD is solving is duller and more familiar to anyone who has wired this up: configure MCP servers for one AI client today and none of that carries over to the next client you install. ARD is an attempt to make registration happen once per organization instead of once per app.

The part the alliance framing skips

All five of these companies are already in a room with Anthropic and OpenAI on this exact subject. Anthropic donated MCP to the Linux Foundation in December 2025 as an anchor project of the new Agentic AI Foundation, alongside Block's goose and OpenAI's AGENTS.md. The foundation's platinum members are AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. MCP is not a competitor's proprietary asset anymore. It is a Linux Foundation project that Google and Microsoft help fund.

At the time of that donation, Anthropic put MCP at roughly 97 million monthly SDK downloads and about 10,000 active servers. That is the head start any discovery layer has to be compatible with, which is exactly why ARD is compatible with it.

What would make this real

A protocol backed by five enterprise vendors is a press release until the manifests exist. The test is cheap and public: fetch https://<vendor>/.well-known/ai-catalog.json and see whether anything is there, then see whether a discovery endpoint returns your agents to a client that did not hardcode them. Snowflake describes its own integration in the conditional ("Snowflake could register the agent"), which is honest and also tells you where this is.

Committee specs from five large companies have a mixed history. This one has an advantage most did not: it is small, it is a static file plus a search endpoint, and it does not ask anyone to rip out what they already run.