Agent Directory Service ADS · /v1/agents

Overview

ADS solves agent discovery at scale using content-addressed storage: each agent record is packaged as an OCI-compliant artifact (reusing container-registry infrastructure and its manifest/layer model) and identified by a content ID (CID) derived from a cryptographic digest ^[1]. Records are indexed into a distributed hash table keyed by a multi-level skill taxonomy (plus domain and module taxonomies), so a query for a capability resolves to the CIDs and peer IDs of agents that claim it, without a central registry ^[1]. Distribution between directory nodes happens over the OCI protocol's peer-to-peer synchronization ^[1].

Distinct from the DHT-based global layer, ADS also runs a host-local "AI Finder" service, exposed over gRPC and, via a gateway, as a REST API: GET /v1/agents lists catalog entries with deterministic pagination, GET /v1/agents/{cid} retrieves one entry, GET /v1/agents/{cid}/export renders a stored OASF record into other ecosystems' native formats (a2a, agent-skill, mcp-ghcopilot), and GET /.well-known/ai-catalog.json serves the host's catalog document at the well-known URI this project explicitly designed to interoperate with the ARD/AI Catalog ecosystem ^[1]. The export operation is described as ADS's "heterogeneity bridge" -- one canonical OASF record materialized into several downstream protocols' native card formats without changing its underlying storage representation ^[1].

ADS does not define its own agent capability schema from scratch; it uses the separately-maintained Open Agentic Schema Framework (OASF, see oasf) as its native record format, and treats other ecosystems' formats (A2A cards, MCP server descriptors) as export targets rather than first-class stored types ^[1].

Current state: the draft was published 6 July 2026 and expires 7 January 2027; it explicitly states it is "to be removed before publishing as an RFC," i.e. still a working draft rather than a finished document ^[1]. Unlike several other records on this map, this one has a concrete, buildable open-source implementation (github.com/agntcy/dir) referenced directly from the spec text, which supports recording implementation as "shipping" -- but no independent count of production deployments or registered agents was located, so adoption remains unmeasured.

What is disputed or unknown: how ADS's DHT-based global layer interacts in practice with a search-first, centralized-registry model like ARD's -- the "AI Catalog Interoperability" section of the draft describes projection and export, but whether third parties beyond AGNTCY's own tooling actually consume ADS-hosted records via the ai-catalog.json path was not independently verified here.

Discovery and probe

Content-addressed OCI-artifact storage plus a DHT for global skill-based discovery; a host-local AI Finder service exposes the same records over gRPC and a REST gateway · /.well-known/ai-catalog.json (host-local AI Finder); GET /v1/agents (REST API)

GET https://{origin}/v1/agents
parse:     200 response is a paginated JSON list of CatalogEntry records; a companion GET /.well-known/ai-catalog.json returns the host's WellKnownCatalog document
on absent: Host does not run an AGNTCY ADS AI Finder REST gateway; the record may still exist in the distributed hash table without a host-local HTTP surface

Live check run by the MCP probe: GET /v1/agents. An Agent Directory Service lists catalog entries as a JSON array at /v1/agents.

{
 "path": "/v1/agents",
 "expect": {
  "content_type": "json",
  "text_match": "\"(agents|entries|items|records)\"\\s*:\\s*\\["
 },
 "note": "An Agent Directory Service lists catalog entries as a JSON array at /v1/agents."
}

Instances

• Agora (agent marketplace) API (api.agoraproto.org) · Observed 2026-10-23

Every instance with its source is listed on the instances page.

Verified SDKs, reference implementations & validators:

• agntcy/dir (AGNTCY Directory reference implementation) · implementation (2026-07-06)

See also

• Builds on: Open Agentic Schema Framework
• Pairs with: Open Agentic Schema Framework, Agentic Resource Discovery (ARD), AI Catalog well-known file, Agent2Agent Protocol (A2A), Model Context Protocol, GB/Z 185 Artificial intelligence — Agent interconnection
• Scanners & registries: AGNTCY Directory specification
• Not to be confused with: ADS is a generic acronym also used for Active Directory Services and Alternate Data Streams; within the agentic space, unrelated to ARDS/ARD

Comparisons

• DNS (A/SRV) & WHOIS vs AID & DNS-AID (classic vs agentic)

References

1. draft-mp-agntcy-ads-02 (6 July 2026) defines ADS as a distributed directory service storing agent metadata as content-addressed OCI artifacts, discoverable via a DHT keyed on a skill taxonomy, and exposes an AI Finder service over gRPC and REST (GET /v1/agents, GET /v1/agents/{cid}, GET /v1/agents/{cid}/export, GET /.well-known/ai-catalog.json) that is explicitly interoperable with the AI Catalog / ARD ecosystem. — https://datatracker.ietf.org/doc/html/draft-mp-agntcy-ads-02 (2026-07-06) VERIFIED
2. AGNTCY published a technical note describing AI Catalog interoperability layered over its directory service. — https://blogs.agntcy.org/technical/2026/06/17/ai-catalog-over-directory.html (2026-06-17) REPORTED

Verification log and errata

• 2026-09-06 — verified against draft-mp-agntcy-ads-02 spec text via datatracker.ietf.org