Opens in a new tabSkip to content
Agent LighthouseAgent Lighthouse

    Searches the text of every published page. The evidence sources themselves are not in this index — search all of them on the trusted sources page.

    GitHub ↗
    Browse checks and page contents
    agent-interfaces/ai-catalog-urls

    AI Catalog entry URLs valid

    What it checks

    Broken service URLs in your AI catalog cause agents to fail when trying to use your services, creating a poor user experience. Verify all URLs are correct, deployed, and returning HTTP 200.

    Why it matters

    A discovery client follows an entry’s url to reach the artifact it advertises. For catalog- or registry-typed entries, it follows that url into a nested catalog. One dead url therefore does not degrade a listing. It truncates a whole branch of discovery.

    Evidence

    ARD entries[].url dereferenced by real clients

    • hf-discover’s navigate() uses an entry’s url to traverse into nested catalogs and federated registries, fetching entries whose type is a catalog or registry media type — github.com/…/hf-discover (verified 2026-08-24)
    • Independent implementations check liveness explicitly: HelgeSverre/ardvark ships internal/crawler and internal/probe, and iFurySt/OpenARD ships internal/cli/verify.go.
    • Live manifests point at operational endpoints an agent would call immediately: Neon’s ten entries are MCP servers and skills, Weaviate’s nine are docs, agent skills, an OpenAPI description and a sitemap, and Shopware’s is a Store-API MCP server url.
    • ARD §4.2 requires exactly one of url or data per entry, so an entry that embeds its artifact has no endpoint to dereference and is fully conformant — github.com/…/ard-spec (verified 2026-08-24)

    Limits

    No vendor document states that a crawler penalises or downranks a site for a dead catalog url. The consequence is mechanical rather than a published ranking signal: traversal stops, and the tool call fails. The federation-following behaviour also lives in a user-driven client. Hugging Face’s hosted server states that “Navigation is intentionally not exposed by the hosted server”. ARD itself is a draft (v0.9). The inline-data case above is also a live false-positive risk that any liveness check must respect. The pre-2026-08-22 implementation was unreachable on real sites for the opposite reason. It aborted unless the manifest exposed a services array, which no spec or deployment uses.

    How it scores

    The dereferencing is done by real, readable code in more than one implementation, which is stronger than a convention. It is not A for two reasons: no vendor documents a penalty for a dead url, and the traversing client is user-driven rather than a hosted crawler.

    Sources