Measured 31 August 2026

Every public ARD registry, compared

There are five public ARD registries. We benchmarked all of them on 31 August 2026 using the specification's own conformance tool. GitHub's Agent Finder fails conformance, one returns nothing for general queries, and none of the four incumbents implements federation: auto, the mode the specification makes the default.

The benchmark

Conformance is the specification's own tool. Precision is top-1 over six queries with unambiguous correct answers. Spread is the average gap between the best and worst score in a result set, which measures whether a ranking carries any information at all.

What the numbers mean

GitHub Agent Finder fails conformance

Its GET /agents endpoint answers 200 with a body that is not a paginated object, and the conformance tool marks the registry FAIL on exactly that. It is otherwise the strongest of the incumbents on result quality, and it is the default finder in the official connectors.

A score spread of 6.7 cannot rank

WellKnown returns accurate top results and holds a large index, but the gap between its best and worst result averages under seven points. A client cannot tell a strong match from a weak one, which defeats the purpose of returning a score.

Hugging Face Discover is not a general registry

It scored zero of six on precision because it indexes only Hugging Face Spaces and Skills. For queries about anything else it returns nothing relevant, despite shipping inside an official CLI.

Nobody implements the default federation mode

This is the largest gap in the ecosystem. The specification makes auto the default: query your peers, merge, return one set. Tested against all four incumbents, auto returned identical results to none, and the only source in every response was the registry's own domain.

Measured across 26 probe queries, the union of all four indexes was 68% larger than the largest single one. That difference is what a client loses every time it has to pick a registry instead of asking one that federates.