Agentic Resource Discovery (ARD) Specification and Hugging Face Implementation

Hugging Face has introduced a reference implementation of the Agentic Resource Discovery (ARD) specification. This open standard enables AI agents to move from a static "install-first" model to a dynamic discovery model, allowing agents to find and integrate tools, skills, and other agents at runtime via a shared discovery layer.

The Agentic Resource Discovery (ARD) Specification

ARD is a draft, open specification developed by a consortium including Microsoft, Google, GoDaddy, and Hugging Face. It provides a discovery layer that sits in front of existing protocols like the Model Context Protocol (MCP) for tool calling, Skills for instruction consumption, and Agent-to-Agent (A2A) for inter-agent communication.

While previous models required developers to hardcode server URLs or dump all tool descriptions into an LLM's context window, ARD moves selection outside the LLM. It uses registries to index capabilities with high-fidelity signals—including publisher identity, compliance attestations, tags, and representative queries—and exposes them via a REST endpoint. This allows agents to perform intent-based search in natural language to find the right capability dynamically.

The specification defines two primary components:

• Static Manifests: A format called ai-catalog.json that allows publishers to host capabilities at a well-known URL.
• Dynamic Registry API: A POST /search endpoint that provides live, ranked discovery of resources.

Hugging Face Discover Tool Implementation

The Hugging Face Discover Tool serves as the reference implementation of ARD. It provides search access to thousands of Skills, ML applications, and MCP Servers hosted on the Hugging Face Hub and other ARD discovery services.

Technical Workflow

Discover leverages the Hub's existing semantic search over Spaces, utilizing an agents=true flag to rank results by agent-oriented metadata. The adapter then applies two primary filters: it only includes Spaces with a RUNNING runtime stage and ensures the response media type matches the request.

Supported media types include:

• application/ai-skill: Generates a SKILL.md file by wrapping the Space's agents.md with required frontmatter (name, description, and source metadata).
• application/mcp-server+json: Provides a catalog entry for Spaces tagged as mcp-server, pointing to the Space's Gradio MCP endpoint over HTTP transport.
• application/vnd.huggingface.space+json: Provides raw Space metadata for clients that wish to handle the data themselves.

Integration and Usage

Users and agents can access the ARD catalog through several interfaces:

Hugging Face CLI

Resource discovery is integrated into the hf CLI. Examples include:

• Searching for training resources: hf discover search "Fine tune a language model"
• Finding MCP servers: hf discover search "Generate an image" --json --kind mcp
• Searching external registries: hf discover search "Purchase aeroplane tickets" --registry-url <catalog-url>

REST API and MCP

Direct access is available via the REST API at https://huggingface-hf-discover.hf.space/search or through the well-known catalog URL at https://huggingface.co/.well-known/ai-catalog.json. Additionally, any MCP client can connect to the search catalog via the MCP endpoint at https://huggingface-hf-discover.hf.space/mcp.

Implications for the Agent Ecosystem

ARD separates discovery from execution, allowing any artifact protocol to use the same envelope without requiring specification-level changes. Because the registry API uses plain HTTP REST, clients can federate against multiple registries. This means a search through one service can surface capabilities hosted by another.

Future updates will include tighter integration with federation modes (auto, referrals, none) and Hub-side support for static ai-catalog.json manifests on user and organization profiles, enabling any Space publisher to advertise their capabilities through standard URIs.