Hugging Face and tech giants launch ARD for dynamic agent tool discovery
New open standard lets AI agents find tools at runtime without manual configuration.
The Agentic Resource Discovery (ARD) specification solves the discovery problem in agent ecosystems. Currently, developers hardcode MCP server URLs or dump all tool descriptions into an LLM's context window—neither scales. ARD introduces a federated registry layer with a static manifest format (ai-catalog.json) and a dynamic REST endpoint (POST /search) that returns ranked capabilities based on natural language queries. It supports rich signals like publisher identity, representative queries, and compliance attestations. This shifts agent capability selection from manual installation and static catalogs to runtime, intent-based discovery.
Hugging Face has implemented ARD in its Discover Tool, a reference implementation that provides search access to thousands of Skills, ML applications, and MCP Servers on the Hub and across other ARD-compliant services. The adapter integrates Hub's semantic search over Spaces, applying filters for running stages and supporting three media types: application/ai-skill (with extracted agents.md), application/mcp-server+json (for MCP-tagged Spaces), and application/vnd.huggingface.space+json (raw metadata). For MCP servers, it generates catalog entries pointing at Gradio MCP endpoints over HTTP. This enables any skill-aware client to discover and invoke capabilities without pre-configuration.
- ARD defines a static manifest (ai-catalog.json) and REST API (POST /search) for federated runtime discovery of tools, skills, and agents.
- Hugging Face's Discover Tool indexes thousands of Spaces, MCP servers, and skills, supporting media types for skills, MCP+s, and raw metadata.
- Replaces install-first model with intent-based search, allowing agents to dynamically find and use capabilities across registries.
Why It Matters
Eliminates manual tool configuration, enabling scalable agent ecosystems with dynamic capability discovery across federated registries.