ARD Resource Discovery

ARD (Agentic Resource Discovery) is an open specification for describing, searching, and discovering AI resources. The official specification credits Junjie Bu (Google), R. V. Guha (Microsoft), and Shaun Smith (Hugging Face) as authors. Its source is available in the official GitHub repository.

Nacos 3.3 exposes published Agents, Skills, Prompts, and MCP Servers to AI applications and toolchains through ARD-compatible endpoints. Clients can search for capabilities and then retrieve their artifacts.

For server setup, authentication, and complete HTTP examples, see the ARD Integration Guide in Ecology. This page explains which resources appear in discovery results and how applications use them.

1. Choose a discovery method

Use case	Entry point
A toolchain needs ARD to search multiple AI resource types, browse catalogs, and download artifacts	ARD Integration Guide
An application needs Agent call definitions, runtime endpoints, and change subscriptions	RAD Integration Guide; RAD stands for Remote Agent Discovery
An application uses Nacos resource APIs directly, or needs to access AgentSpecs	Consult the Client API for the relevant resource type

ARD currently covers Agents, Skills, Prompts, and MCP Servers. AgentSpecs are accessed through native Nacos APIs and are not included in ARD results.

2. Make resources discoverable

Create and publish resources through their management pages. No separate ARD registration is required:

Follow the Agent, Skill, Prompt, or MCP guide to create a resource. Provide a name, description, and business tags to help users search and filter.
Submit and publish it through the AI Resource Lifecycle. Confirm that the resource is enabled and latest points to an online version. Drafts and reviewed but unpublished versions do not appear in discovery results.
Check visibility and caller permissions. The built-in visibility plugin allows non-owners to read PUBLIC resources. PRIVATE resources are visible to the owner, administrators, or identities with explicit visibility grants. API authentication and authorization still apply. Anonymous callers need anonymous access enabled and the required API permissions, and can discover only public resources.
Query the catalog or search as the integration account, and check the resource name, version, and URL.

With the built-in policy, new Agents and MCP Servers default to PUBLIC, while Skills and Prompts default to PRIVATE. For a private release, set scope before submitting for publication. New versions do not reset existing private settings. See Visibility Plugin for defaults, custom-plugin behavior, and scope changes.

The public namespace and PUBLIC visibility scope are separate settings; the namespace alone does not determine whether a resource is public.

Search returns the current online version selected by latest, rather than listing every historical version as a separate result. Search synchronization may take time after a resource is published, taken offline, or updated. Results may also be temporarily incomplete after startup or upgrade. When a resource is missing, check its publishing state and permissions before retrying.

3. Filter and use results

ARD type is the media type of the downloadable artifact, rather than the Nacos resource type name. Use query.filter.type to select formats your application can handle:

Resource	`type`	How to use the artifact
Agent: A2A representation	`application/a2a-agent-card+json`	Parse the AgentCard with an A2A client.
Agent: Nacos definition	`application/vnd.nacos.ai-agent+json`	Read the versioned Agent definition and call interfaces; use RAD for runtime endpoint discovery and subscriptions.
Skill	`application/agent-skills+zip`	Retrieve the ZIP containing `SKILL.md` and packaged resources for a tool that supports Skills.
Prompt	`application/vnd.nacos.ai-prompt+json`	Read the Prompt content and variable definitions, then use the template with application inputs.
MCP Server	`application/mcp-server-card+json`	Read the MCP Server description and connect to the actual service with an MCP client.

Read identifier, version, type, and url from a result, then download the artifact from url. Search responses place resources in results, list responses in items, and catalog responses in entries. See the ARD Integration Guide for pagination and download examples.

3.1. Agent representations

A single ARD query returns at most one entry per Agent. If the current latest contains only A2A interfaces and can export a complete AgentCard, the default is the A2A representation. Definitions containing multiple protocols prefer the Nacos representation. Applications that support only A2A should filter by the A2A media type.

Availability of the A2A representation depends on the current latest version. An older version supporting A2A does not cause its AgentCard to be returned for the current version in ARD.

An Agent artifact contains the definition for a specific version, without runtime endpoints or live health state. Default health flags on declared addresses do not represent health probes either. To select healthy running instances, use RAD to discover endpoints, then invoke them with the appropriate protocol client.

3.2. Vector plugins and resource discovery

ARD can use the AI Vector Plugin to retrieve similar content and rank it together with keyword results. Clients do not need to change the ARD request format. See the plugin guide for server configuration, index verification, and custom Provider integration.

Keyword search remains available without a usable vector plugin. The default embedding method uses local hashing; evaluate relevance against resource descriptions and representative queries. Vector retrieval does not expand visibility or bypass publication requirements.

4. Verify the integration

Search with the application’s identity and namespace, and confirm that the published resource is found.
Fetch the result’s url and check that the version and Content-Type match what the application expects.
Verify Skill and Prompt loading in the tool or application. For Agents and MCP Servers, also verify connectivity and credentials for the actual service.

ARD provides search, catalogs, and artifact retrieval. Edit, publish, and take resources online or offline through the Nacos console, Admin API, or Maintainer SDK.