ARD: Big Tech's Agent Standard vs Anthropic's MCP

The next front in the AI wars isn’t a bigger model — it’s the plumbing. Around July 13, 2026, a coalition of the biggest names in enterprise software lined up behind a new open standard called Agentic Resource Discovery (ARD), a specification for how AI agents automatically find and connect to the tools, data, and other agents scattered across a company’s software stack. The backers read like a who’s-who of the enterprise: Google, Microsoft, Salesforce, Snowflake, and ServiceNow, joined by Cisco, Databricks, GitHub, Hugging Face, and NVIDIA.

Two names are conspicuously absent from the list: Anthropic and OpenAI — the two labs whose standards ARD is, in effect, designed to route around. Reporting framed the coalition as an attempt by the traditional software giants to “beat back” the frontier labs in the fight over enterprise agent infrastructure. It is a standards battle, and standards battles are how incumbents encircle upstarts.

What ARD actually does

To understand ARD, start with the layer it targets. Over the past 18 months, Anthropic’s Model Context Protocol has become the de facto way to connect an AI agent to external tools — the “USB-C for AI,” as its boosters like to call it. MCP standardizes the connection between a model and a single tool or data source: how the agent calls it, what it gets back, how authentication flows.

ARD aims one layer up. Where MCP answers “how does an agent talk to a tool it already knows about,” ARD answers “how does an agent discover which tools exist in the first place.” The specification centers on ai-catalog.json manifests — machine-readable files that let an agent query a centralized registry and enumerate what tools and other agents are available for a given task, across a sprawling corporate software estate. Think of it as a directory service for agents: less a replacement for MCP than a discovery layer that sits above whatever connection protocol does the actual work.

That distinction matters, because it lets the coalition frame ARD as complementary rather than combative — a standard for finding resources, not for talking to them. But the strategic effect is the same. Whoever owns the discovery layer owns the map of the enterprise, and the map is where the leverage sits.

ARD vs MCP: what’s the difference?

The two standards are easiest to keep straight as layers of the same stack: MCP standardizes the conversation between an agent and a tool it already knows about, while ARD standardizes the phone book the agent consults to find that tool in the first place.

	ARD	MCP
What it standardizes	How agents discover which tools, data, and agents exist across a company’s stack	How an agent connects to a specific tool or data source — calls, responses, auth
Core artifact	ai-catalog.json manifests queried through a centralized registry	Client–server protocol; tools exposed as MCP servers
Backed by	Google, Microsoft, Salesforce, Snowflake, ServiceNow, plus Cisco, Databricks, GitHub, Hugging Face, and NVIDIA	Created by Anthropic; native support from Anthropic, OpenAI, Google, and Microsoft
Governance	New coalition-led open specification	Agentic AI Foundation (AAIF) under the Linux Foundation, since December 2025
Maturity	Announced July 2026; adoption untested	De facto standard — tens of millions of monthly SDK downloads, thousands of public servers

On paper, then, “ARD vs MCP” is a false fight: an agent could use ARD to find a tool and MCP to talk to it, and the coalition itself pitches the standards as complementary. In practice they are competing for strategic ground, because the layer enterprises standardize on determines where agent infrastructure — and the sticky revenue attached to it — ends up living.

Why the incumbents are moving now

The competitive logic is straightforward once you look at who’s on the list. Salesforce, Snowflake, and ServiceNow collectively sit on top of a huge share of the world’s enterprise data and workflows. Google and Microsoft own the clouds that data runs on. A discovery standard co-signed by those five hands enterprises a credible alternative to building their agent stacks on a protocol governed by a competitor — and it does so without any single vendor having to win outright.

The subtext is a different vision of what enterprise AI looks like. OpenAI and Anthropic have largely built their enterprise strategies around chatbot-centric products — a conversational assistant as the front door to everything, an approach visible in OpenAI’s ChatGPT Work push. ARD’s backers are betting on something else: agents that operate autonomously inside existing software, discovering and invoking tools without a chat window as the bottleneck. In that world, the valuable real estate isn’t the model or the chat interface — it’s the registry of resources the agents are wired into.

The awkward part: everyone cooperates and competes at once

Here’s where the enterprise AI landscape gets genuinely tangled. Many of the same companies backing ARD also participate in the Agentic AI Foundation (AAIF) — the Linux Foundation directed fund that now governs MCP after Anthropic donated the protocol to it in December 2025. MCP itself has become enormous: on the order of tens of millions of monthly SDK downloads, thousands of public servers, and native support from Anthropic, OpenAI, Google, and Microsoft alike.

So the picture is not two rival camps drawing clean battle lines. It’s the same set of giants cooperating inside a shared foundation while simultaneously backing a competing standard in the market — hedging every bet, supporting every protocol, and making sure that whichever layer wins, they have a seat at the table. Google, for its part, already pushes its own Agent-to-Agent (A2A) protocol for agents to coordinate with each other, a third acronym in an increasingly crowded stack of MCP, A2A, and now ARD.

For enterprises trying to build production agent systems, that proliferation is a double-edged sword. More open standards mean less lock-in to any single vendor — but a stack with three overlapping protocols, each backed by a different coalition, is also a stack that can fragment. The promise of interoperability only pays off if the standards actually converge.

What this signals about the market

Standards fights happen when a market matures past the land-grab phase and the incumbents realize the terms of competition are being set by someone else. That MCP became dominant enough to trigger a counter-coalition is itself a measure of how far Anthropic’s protocol has spread — you don’t organize a five-company alliance against a standard nobody uses.

It also reflects where the value in agentic AI is migrating. As raw model capability converges — a trend visible in everything from the state of AI coding assistants to meta-harnesses for orchestrating agents — the durable advantage shifts from the model itself to the infrastructure around it: the data it can reach, the tools it can call, the registry that tells it what exists. The frontier labs built the best models; the enterprise incumbents own the data and workflows. ARD is the incumbents pressing that advantage.

What it means

ARD is less a product launch than a positioning move — a signal that the enterprise software establishment intends to shape the agent-infrastructure layer rather than inherit whatever the frontier labs define. Whether developers actually adopt ai-catalog.json manifests at scale is unknown; standards succeed on adoption, not press releases, and MCP’s head start is real.

Who wins: the enterprise incumbents, who get a shared standard that keeps agent infrastructure anchored to their platforms rather than a competitor’s; and enterprise buyers, who gain leverage from having a credible alternative to any single vendor’s stack. Interoperability, in principle, wins too.

Who’s pressured: Anthropic and OpenAI, whose de facto ownership of the connection and chat layers is exactly what ARD is designed to contest. Neither is threatened at the model frontier, where they still lead — but the plumbing is where recurring, sticky enterprise revenue accrues, and that’s the ground being fought over.

What to watch: whether ARD gains real developer traction or stalls as a coalition announcement; whether the overlapping standards — MCP, A2A, ARD — converge or fragment the stack; and whether Anthropic and OpenAI respond by extending their own protocols upward into discovery. The models get the headlines, but in enterprise AI, the winner is increasingly decided by who controls the layer beneath them.