In the rapidly evolving world of AI agents, enabling them to discover and utilize the right tools, skills, or even other agents is a monumental challenge. Current methods often involve hardcoding connections or manually integrating capabilities, a process that quickly becomes unwieldy. Imagine an agent that could intelligently search for and integrate new functionalities on demand, without needing prior setup.
This is where Agentic Resource Discovery (ARD) steps in, providing a crucial missing layer in the agent ecosystem. It’s a groundbreaking, open specification designed to empower AI agents with dynamic, intent-based discovery capabilities. Developed collaboratively by industry leaders like Microsoft, Google, GoDaddy, and Hugging Face, ARD is poised to revolutionize how AI agents interact with the vast universe of available tools.
The Challenge of Agent Tool Discovery
Today, if you’re building with AI agents, you’re likely familiar with foundational protocols like MCP (Multi-Agent Communication Protocol) for tool calls, Skills for consuming instructions, and A2A for agent-to-agent communication. While these protocols provide the mechanisms for agents to *use* capabilities, they all share a common limitation: the user or developer is still responsible for *knowing*, *discovering*, and *integrating* those capabilities beforehand. This “install-first, use-later” model works for a handful of frequently used tools but hits a significant roadblock when trying to scale to thousands of diverse, ad-hoc services.
The common fallback—dumping every available tool description into the Large Language Model’s (LLM) context window—is equally problematic. This approach is severely constrained by the LLM’s context budget, making it impractical for a truly extensive range of options. Furthermore, many tool descriptions are too thin to allow the LLM to accurately disambiguate between similar functionalities, leading to suboptimal choices.
Introducing Agentic Resource Discovery (ARD)
ARD directly addresses these challenges by shifting the capability selection process outside the LLM’s direct context window. Instead of pre-installing or context-stuffing, ARD defines a standard for how agents and tools are cataloged, indexed, and searched across federated registries. This means an agent can find and integrate new capabilities dynamically at runtime, without needing them pre-configured.
Imagine a global directory for AI tools and services. An ARD registry indexes capabilities with rich metadata, including publisher identity, representative queries, compliance attestations, and descriptive tags. When an agent needs a specific function, it can issue a natural language query to an ARD endpoint, and the registry intelligently returns the most relevant capabilities. This powerful shift from static catalogs to intent-based search enables agents to tap into an ever-growing ecosystem of MCP tools, A2A agents, and other services on the fly.
Hugging Face’s Pioneering ARD Implementation
Hugging Face has launched its Discover Tool, serving as a powerful reference implementation of the ARD specification. This innovative tool provides unparalleled search access to thousands of Skills, Machine Learning applications, and MCP Servers, not just on Hugging Face but across other ARD-compliant discovery services. It’s a tangible example of how ARD brings dynamic capability discovery to life.
The Discover Tool ingeniously combines the Hugging Face Hub’s existing semantic search capabilities over Spaces with its Agent Skills. It then serves these results as standardized ARD catalog entries. The Hub already hosts a vast catalog of Spaces running Gradio apps, MCP servers, and demos, and its semantic search can be filtered with an agents=true flag to prioritize agent-oriented metadata. The Discover Tool seamlessly translates these searches into the ARD specification, making it a powerful gateway to AI capabilities.
When an agent requests a capability, the Discover Tool applies smart filtering. It ensures that only Spaces currently in a RUNNING state are returned, guaranteeing active and usable resources. Moreover, the response media type is dynamically driven by the request, supporting specific formats:
application/ai-skill: For discovering and integrating agent skills.application/mcp-server-card+json: For finding MCP servers.application/ai-application+json: For general AI applications.
For skills, the adapter reads any agents.md file within a Space, wrapping its description with the necessary metadata (name, description, source) for a skill consumer. For MCP-tagged Spaces, it generates a catalog entry pointing directly to the Space’s Gradio MCP endpoint, making it immediately accessible to any MCP client.
Getting Started with ARD and the Discover Tool
Accessing the power of ARD through the Hugging Face Discover Tool is straightforward, whether you prefer command-line interaction, a REST API, or an MCP client. This flexibility ensures developers can integrate ARD discovery into their workflows with ease.
To get started with the Hugging Face CLI, simply install it and begin searching:
- Install the CLI:
uv tool install huggingface_hub - Search for resources to train a model:
hf discover search "Fine tune a language model" - Find MCP Servers to generate an image:
hf discover search "Generate an image" --json --kind mcp - Search other registries by specifying a URL:
hf discover search "Purchase aeroplane tickets" --registry-url <catalog-url>
For direct access, the Hugging Face catalog is published at its well-known URL: https://huggingface.co/.well-known/ai-catalog.json. You can query the Discover Tool’s search endpoint directly using a REST API:
POST https://huggingface-hf-discover.hf.space/search
Example using curl to find skills:
curl -s https://huggingface-hf-discover.hf.space/search \
-H "Content-Type: application/json" \
-d '{ "query": { "text": "fine tune a sentence transformer", "filter": { "type": ["application/ai-skill"] } }, "pageSize": 5 }'Alternatively, you can connect any MCP Client to search via the MCP endpoint at https://huggingface-hf-discover.hf.space/mcp to dynamically discover tools.
The Future of Agentic AI
The Agentic Resource Discovery specification represents a significant leap forward by fundamentally separating discovery from execution. Its design promotes incredible flexibility: a static manifest format driven by media types means any artifact protocol can leverage the same discovery envelope without requiring specification-level changes. The registry API is a simple HTTP REST interface, allowing any client to easily federate against it.
The Hugging Face Discover Tool is more than just an implementation; it’s a living testbed for this design. It demonstrates that ARD doesn’t require inventing new artifact formats. Instead, it wraps existing search backends, like the Hub, in the specification’s envelope, allowing the same underlying resources to surface as skills or MCP servers depending on the client’s request. This robust federation is a cornerstone of ARD, enabling a search through one service to seamlessly uncover capabilities hosted by another.
Future developments include tighter integration with ARD’s federation modes (auto, referrals, none) and Hub-side support for static ai-catalog.json manifests on user and organization profiles. Once implemented, any Space publisher will be able to advertise their capabilities through the standard well-known URI mechanism, fostering an even more open and discoverable ecosystem. ARD is not just a specification; it’s the blueprint for a truly dynamic and interconnected future for AI agents.
Source: Hugging Face Blog
