On June 17, 2026, eleven companies, Google, Microsoft, Salesforce, Snowflake, ServiceNow, Cisco, Databricks, GitHub, GoDaddy, Hugging Face, and Nvidia, published a new specification for how AI agents find the tools, APIs, and data sources they need. Two names are missing from that list: Anthropic and OpenAI. Given that Anthropic's Model Context Protocol has become the closest thing the agent era has to a universal standard, the absence reads less like an oversight and more like a message.

11
Companies backing the new Agentic Resource Discovery specification
97M+
Monthly SDK downloads for Anthropic's Model Context Protocol as of December 2025
150
Organizations running Google's competing A2A protocol in production, not pilot

What ARD does

The new specification is called Agentic Resource Discovery, or ARD. Its mechanism is duller than the politics around it. Each participating company publishes a manifest file, ai-catalog.json, at a well-known path on its domain, listing the tools, APIs, MCP servers, and agents it exposes. Federated registries crawl those manifests and answer plain-language discovery queries from agents looking for the right tool to complete a task. The specification is Apache 2.0 licensed and built on a data model maintained by a Linux Foundation working group.

Importantly, ARD does not replace Anthropic's MCP. It sits above it, as a discovery layer, and is explicitly designed to complement existing protocols rather than compete with them directly, including MCP and Google's Agent2Agent protocol. An agent still needs MCP, or something like it, to connect to a tool once it's found. ARD only changes how that tool gets discovered in the first place.

Why the backer list is the real story

MCP has effectively won the agent-to-tool layer. Anthropic open sourced it in November 2024, donated it to the Linux Foundation's Agentic AI Foundation in December 2025, and it now sits at more than 97 million monthly SDK downloads with adoption across every major lab, OpenAI included. That's the problem for the companies whose software holds most of the enterprise data agents need to reach. None of them want the connective tissue of the agent era defined entirely by a protocol one competitor built and still leads.

The nuance worth not glossing over: This isn't a clean split between an Anthropic camp and a rival camp. The Agentic AI Foundation that governs MCP was co-founded by Anthropic, Block, and OpenAI, with Google, Microsoft, AWS, Cloudflare, and Bloomberg among its supporters, several of the same companies now backing ARD. These companies are participating in MCP's governance and simultaneously building a layer they control on top of it. That's a more accurate picture than a straightforward standards war: less "beat Anthropic," more "don't let Anthropic's design be the only door."

The stack is getting crowded, not simpler

ARD is the newest addition to a stack that already has several layers doing related but distinct jobs. MCP connects an agent to a specific tool or data source. Google's Agent2Agent protocol, now in production at 150 organizations rather than pilot, handles how separate agents communicate and delegate work across organizational boundaries. WebMCP, an emerging layer with reported support already in Chrome Canary, renders interactive tool interfaces directly inside a chat window. ARD now sits above all of it, answering the question of how an agent finds out what's available before any of the other layers get involved.

Each layer is individually reasonable. Collectively, a business building agent integrations today has to track four overlapping specifications, backed by different, partially overlapping coalitions, at different stages of maturity, to understand what it's building on.

What this means for a business building automations now

A fragmented standards layer is a mixed outcome for buyers. A genuine challenge to a single default protocol keeps that layer competitive rather than owned outright by one company. The cost is that agent-to-tool integrations built this year could need rework depending on which layers win adoption and which quietly fade. The defensive posture is the same one that's mattered in every prior platform standards fight: keep integrations abstracted behind a layer you control, rather than hard-wired to any one vendor's specification, so a shift in which standard wins doesn't mean rebuilding the automation itself.

Sources: AI Pulse · Big Picture · workplaceai.ai. ARD specification, its backers, and mechanism: AI Weekly, "Google, Microsoft back ARD spec to counter Anthropic's MCP," and Crypto Briefing, "Google, Microsoft, and Salesforce back new AI software standard to counter OpenAI and Anthropic," June 20, 2026. The "fight moved from the model to the standard" framing and the buyer guidance on abstraction: InfoTech Research Group, "Big 5 AI Vendor Roundup: Week of July 13, 2026." MCP's download figures, its Linux Foundation donation, and the Agentic AI Foundation's founding members: ServiceNow Community, "MCP: The Protocol Powering Agentic AI," and AI Weekly's MCP coverage archive. Google's A2A protocol adoption figures and production status: The Next Web, "Google Cloud Next 2026," May 6, 2026. Every figure above is attributed to its original reporting; none is a WorkplaceAI study.