Microsoft and Google: New Standard for AI Agents

A New Era for AI Agent Discovery

In the field of intelligent agent development, three main protocols currently dominate: MCP, which provides agents with a standard way to interact with tools; Skills, which allow agents to follow specific instructions; and A2A, which facilitates communication between agents. However, these protocols rely on the idea that the user already knows which tool or agent to use, placing the burden of discovery and integration of these capabilities on them.

The Agentic Resource Discovery (ARD) specification emerges as a solution to this problem. Developed by contributors from Microsoft, Google, GoDaddy, Hugging Face, and others, ARD aims to create an open standard for cataloging, indexing, and searching for tools and agents across federated registries. This would enable an agent to dynamically discover capabilities during execution, without requiring prior installation. This standard is not a commercial product but a framework that any company can adopt to enrich its ecosystem of agents and tools.

Challenges of Current Discovery

Today, the model for discovering agent capabilities relies on an install first, use later approach. Developers must hard-code the URLs of MCP servers into configuration files, and users must manually connect services to their AI applications. This method works for a limited number of tools used daily, but it is not suitable for a larger scale where thousands of ad hoc services might be needed.

To circumvent these limitations, some have turned to solutions like injecting tool descriptions into the context window of the LLM. However, this approach is limited by the available context capacity. Other search-based strategies exist, but they often suffer from overly vague descriptions to be truly useful.

ARD proposes to move the selection outside of the LLM. A registry indexes capabilities with detailed information such as the publisher's identity, sample queries, compliance attestations, and tags. It exposes a REST endpoint to allow natural language search, enabling the model to invoke the found capabilities. This shift from a static catalog to dynamic search allows agents to discover and access a wide range of tools and services without preconfiguration.

The ARD specification relies on two key elements:

• A static manifest format named ai-catalog.json, allowing publishers to host their capabilities at a well-defined URL.
• A dynamic registry API at POST /search, offering real-time and ranked discovery.

Implementation by Hugging Face

Hugging Face Discover is the reference implementation of the ARD specification. This tool provides search access to thousands of Skills, machine learning applications, and MCP servers, not only on Hugging Face but also across other ARD discovery services.

Discover operates by combining semantic search from the Hub on Spaces with Agent Skills, serving the results as entries in the ARD catalog. The Hub already hosts a catalog of Spaces running Gradio applications, MCP servers, and demos. Its semantic search includes a flag agents=true that returns Spaces ranked by agent-oriented metadata, and Discover translates this search into the ARD specification.

The Discover adapter applies two main filters. First, it only returns Spaces whose execution state is RUNNING. Second, the media type of the response is determined by the query, with three media types supported:

• application/ai-skill: the default type, wrapping the agents.md file of the Space.
• application/mcp-server+json: a catalog entry for Spaces labeled mcp-server.
• application/vnd.huggingface.space+json: raw metadata of the Space for client processing.

For Spaces labeled MCP, the adapter generates a catalog entry pointing to the Space's Gradio MCP endpoint via HTTP. The URL uses the execution domain of the Space, or the default convention .hf.space.

Discover is integrated into the Hugging Face CLI (hf). To get started and provide access to you or your agent, you can run the following commands:

• hf discover search "Fine tune a language model"
• hf discover search "Generate an image" --json --kind mcp
• hf discover search "Purchase aeroplane tickets" --registry-url <catalog-url>

Using the REST API and MCP Servers

It is also possible to search directly in the catalog using the REST API or an MCP server.

The Hugging Face catalog is accessible at the following URL:

https://huggingface.co/.well-known/ai-catalog.json

To perform a direct search, you can use the following command:

POST https://huggingface-hf-discover.hf.space/search
curl -s https://huggingface-hf-discover.hf.space/search \
-H "Content-Type: application/json" \
"text": "fine tune a sentence transformer",
"type": ["application/ai-skill"]

Searching for MCP Servers

To search for MCP servers, use the following command:

curl -s https://huggingface-hf-discover.hf.space/search \
-H "Content-Type: application/json" \
"text": "transcribe some audio",
"type": ["application/mcp-server-card+json"]

Alternatively, you can connect any MCP client to perform a search via the MCP endpoint at https://huggingface-hf-discover.hf.space/mcp.

Implications for the ARD Specification

The ARD specification separates discovery from the execution of capabilities. The static manifest format is defined by the media type, allowing any artifact protocol to use the same framework without modifications. The registry API is based on HTTP REST, enabling any client to federate with it. Discover is one of many reference implementations of the specification, and thanks to the federation built into the protocol, a search across one service can reveal capabilities hosted by another.

The Discover tool serves as a functional test for this design. It does not invent a new artifact format but wraps an existing search backend, the Hub, within the framework of the specification, allowing the same Spaces to present themselves as skills or MCP servers as needed by the client.

Next steps include closer integration with the federation modes of the specification (auto, references, none) and Hub-side support for static manifests ai-catalog.json on user and organization profiles. Once these elements are in place, any Space publisher will be able to promote their capabilities via the well-known URI mechanism.