エージェントがMCPサーバーを実行時に探す標準ARD、Googleら11社が公開 エージェントがMCPサーバーを実行時に探す標準ARD、Googleら11社が公開

A group of eleven companies, including Google, has published ARD, a specification intended to let AI agents discover Model Context Protocol (MCP) servers at runtime rather than requiring every connection to be configured in advance. The proposal matters because the "connect everything first" approach that works fine for a couple of servers becomes unworkable once an organization runs many agents and tools at the same time.

MCP, which Anthropic introduced in late 2024, has quickly become a common way to link language-model agents to external tools, data sources, and services. An MCP server publishes a set of capabilities — typically tools, resources, and prompts — that a client agent can invoke. The usual workflow today is to write each server's endpoint and credentials into a configuration file before the agent starts. For one or two servers, this is trivial. But in a setting with dozens of internal MCP servers and agents, plus a long tail of external tools, pre-wiring all of them by hand breaks down. Configuration files grow stale, every new tool requires a manual edit, and agents end up holding connections they may never actually use.

ARD reframes this as a discovery problem. Instead of embedding a fixed list of servers, an agent would look up what is available at the moment it needs something. Based on the published description, the standard appears to define a consistent way to advertise and locate MCP servers — for example, through a well-known location tied to a domain or a queryable catalog — so that an agent can resolve which servers exist, what they offer, and how to connect to them. This mirrors long-standing patterns in distributed systems, where service discovery mechanisms such as DNS-based service discovery or registries like Consul let clients find services dynamically rather than hard-coding addresses. The use of conventional well-known URI patterns, formalized in RFC 8615, is a likely building block for that kind of lookup.

The involvement of multiple vendors is notable. A discovery standard is only useful if many clients and servers implement it the same way, so a multi-company effort improves the odds of interoperability compared with a single-vendor scheme. It also fits a broader industry movement toward standardized metadata that agents can read on their own. Anthropic has been building an official MCP registry, a directory meant to help users and tools find published servers. Google's Agent2Agent (A2A) protocol uses "agent cards," machine-readable documents at a known URL that describe an agent's identity and capabilities so other agents can decide whether and how to interact. ARD appears to extend that discovery-metadata thinking specifically to the MCP server layer.

Several practical questions sit alongside discovery. Finding a server is one thing; deciding whether to trust it, authenticate to it, and authorize its actions is another. MCP has been evolving its own authorization story, including OAuth-based flows, and runtime discovery raises the stakes because an agent may encounter servers it has never seen before. Governance, identity, and the risk of malicious or spoofed endpoints will likely shape how cautiously organizations enable automatic discovery, particularly across trust boundaries between internal systems and the public internet. A standard can describe how to locate and describe servers, but operators still need policies for which ones an agent is permitted to use.

It is worth noting that this report originates from a developer blog post rather than a formal specification release, so details such as the exact endpoint format, the metadata schema, and the list of all eleven participating companies should be confirmed against primary sources before being relied upon. The direction, however, is consistent with where the ecosystem has been heading: