The Layer Above MCP: Eleven Companies Line Up Behind an Agent Standard Anthropic Didn't Write

A standards fight nobody announced as a fight

The most consequential AI story of the past two days is not a model launch. It is a JSON file.

On July 13, The Information reported that Google, Microsoft, Salesforce, Snowflake and ServiceNow had agreed to support a shared technical standard for connecting AI agents to business software — a move that aggregators immediately framed as a counterstrike against Anthropic's Model Context Protocol. (The Information's piece sits behind a paywall; the framing here comes from secondary coverage, and I could not read the original reporting.) The standard in question is real and public, though, and it has a name: Agentic Resource Discovery, or ARD.

ARD is not new as of this week. It landed on June 17, 2026 as a v0.9 draft, licensed Apache 2.0, announced on the Google Developers Blog by engineers Junjie Bu and Srinivas Krishnan. What is new is the political weight now stacked behind it. Per Search Engine Journal, eleven organizations have signed on: Google, Microsoft, GitHub, Hugging Face, Cisco, Databricks, GoDaddy, NVIDIA, Salesforce, ServiceNow and Snowflake.

Read that list again for who is missing. OpenAI is not on it. Anthropic — the company that wrote MCP — is not on it either.

What the spec actually does

The technical idea is narrow and, honestly, sensible.

Today, an agent can only use tools someone wired into it in advance. You install an MCP server, you configure a connection, and the tool definitions get stuffed into the model's context window. That works at ten tools. It falls apart at ten thousand, because the catalog stops fitting in the prompt.

ARD's answer is to move tool selection out of the context window and into a search service. It rests on two primitives. First, a publisher hosts a machine-readable manifest — typically ai-catalog.json — at a well-known path on its own domain, listing the capabilities that domain offers. Second, registries crawl those catalogs, index them, and expose a REST POST /search endpoint that returns ranked, live results. An agent describes what it is trying to accomplish in natural language; the registry tells it which tool to call. Domain ownership verification is meant to establish that the tool is what it claims to be. Google's post frames the whole thing as answering three questions agents currently cannot answer in a standard way: "Where does the right capability live? Which capability should I actually use? And how do I verify it's safe to connect to?"

Hugging Face's launch post puts the shift more bluntly: "from manually installed, static catalogs to intent-based search." HF has already shipped a reference implementation — a Discover Tool that exposes thousands of Skills, ML applications and MCP servers through the search interface.

The part that complicates the "counterstrike" story

Here is where the day's framing and the day's documents disagree, and the gap is worth sitting in.

Every primary source I read describes ARD as a layer above MCP, not a replacement for it. Google's announcement is explicit that the spec is protocol-agnostic and that MCP servers and A2A agents are example capability types you publish inside a catalog; after discovery, the agent talks to the tool in that tool's native protocol. The spec even defines a media type, application/mcp-server-card+json, whose entire purpose is to make MCP servers surfaceable through ARD search. Hugging Face positions ARD as sitting on top of a three-protocol stack — MCP for tool calls, Skills for instructions, A2A for agent-to-agent — and insists: "It is not a product or a marketplace. It is a shared standard that any company can implement independently."

Snowflake's post describes a problem that is pure plumbing, not politics: "Today, configuring MCP connections for one AI client does not automatically make those agents available in the next one." Their pitch is that an agent a data team ships would become discoverable across Snowflake CoWork, Claude, Copilot or a homegrown app without being re-registered in each. Note that Claude is on that list. This is not a spec designed to route around Anthropic; on its face it makes Anthropic's protocol more findable.

So which is it — an ambush or an accessory?

Why "complementary" and "strategic" are not opposites

The honest answer is that both readings can be true, and the tension is the story.

MCP won the invocation layer. Per the coverage of The Information's report, it has quietly become the default over roughly the past eighteen months, and it did so on a foundation owned by a competitor. That is an uncomfortable position for Microsoft, Google and Salesforce, and no amount of Apache 2.0 licensing makes it comfortable. Discovery is the layer above invocation — the layer that decides which tool an agent even considers. Whoever runs the registries shapes what agents see, in the same way search engines shape what humans see. Owning the index has historically been worth more than owning the pipe.

And look at who signed. As Crypto Briefing notes, this coalition runs the systems where enterprise data actually lives — Salesforce for customers, Snowflake for data, ServiceNow for workflows, plus the two largest clouds, plus NVIDIA and GitHub. If they publish ai-catalog.json and OpenAI and Anthropic do not build against it, the model companies become tenants in a discovery graph the software incumbents own. Complementary to MCP, strategically adjacent to Anthropic. Both.

One more wrinkle cuts against the conspiracy read: the underlying AI Catalog data model is maintained by a Linux Foundation working group, and per the reporting, OpenAI and Anthropic are members of the broader Linux Foundation agent-standards effort. This is a fight being conducted almost entirely inside a shared room.

Hype versus real

Real: the spec exists, is public, is versioned, and has running code. GitHub's Copilot agent finder reportedly supports discovery across MCP services. Hugging Face's registry is live. Eleven companies with genuine distribution have their names on it.

Not yet real: adoption. This is a v0.9 draft. Its value is entirely a function of whether a registry ecosystem materializes, and Search Engine Journal notes that ecosystem remains at an early stage. A discovery standard with no crawlers and no catalogs is a schema, not an infrastructure. The web has a long graveyard of well-designed manifest files that nobody published — Google's own John Mueller has publicly argued that LLM systems cannot meaningfully distinguish sites that ship machine-readable files like llms.txt, which is a pointed thing for a Google employee to have said about this genre of idea.

Also unconfirmed: any statement from Anthropic or OpenAI about ARD. Their absence from the supporter list is a fact. What that absence means — refusal, indifference, or a conversation still in progress — is not something the sources establish, and I am not going to pretend otherwise.

The takeaway

Treat the headline framing with suspicion and the underlying move with respect. The claim that eleven companies have banded together to kill MCP does not survive contact with the specification, which is built to make MCP servers easier to find. But the claim that this is merely a technical convenience does not survive contact with the guest list.

The pattern to watch is not whether ARD "beats" MCP. It is whether discovery becomes the layer where leverage in agentic AI actually accumulates — and if it does, whether the companies that own the enterprise data estate end up owning the index that every agent, including Claude and ChatGPT, has to query to find its way around. That would not be a defeat for MCP. It would be a demotion: from the standard everyone builds on, to a payload format inside somebody else's catalog.

Ask again in six months how many ai-catalog.json files exist in the wild. That number, not this week's press coverage, is the whole story.