Add Desvela ARD Registry to agent-finders.json

[mrenzo20]

Adds a third Agent Finder to the default menu: the Desvela ARD Registry.

Both finders that ship today (GitHub, Hugging Face) index their own ecosystem. This one is a general crawl of the open web: /.well-known/ai-catalog.json, /llms.txt and /agents.md across the Tranco top-100K, refreshed weekly, so it surfaces publishers that are not on either platform.

It implements the interaction contract in shared/discovery-instructions.md as-is:

• POST /search with {"query": {"text": "..."}}, and filter.type narrowing works
• results carry identifier, displayName, type, url, description, score, source
• score is relevance only, never a trust or safety rating
• remote MCP at POST /mcp (Streamable HTTP), which also exposes a free preflight(domain) tool
• pageToken pagination

Verified against the spec's own conformance suite (conformance/bin/conformance-test registry https://registry.desvela.dev): PASS, 0 critical errors. /explore (§7.3) is implemented as well.

Happy to adjust naming or the description if you'd rather word it differently.

Disclosure: I run this registry.

[mrenzo20]

Some context up front, since I saw the same question on #2 and Desvela is not a name anyone here has reason to know.

What it indexes. A weekly crawl of the Tranco top-100K, four surfaces per domain: /.well-known/ai-catalog.json, /llms.txt, /agents.md, and the AI bot directives in robots.txt. There is no platform behind it, so it surfaces publishers that are not on GitHub or Hugging Face. Where that crawl stands today: 8 domains in the top-100K serve a valid ai-catalog.json with at least one entry, 5,445 serve llms.txt, 888 serve agents.md. The index holds 79 live catalog entries. Those numbers are small because adoption on the open web is small, and being able to state them at all is most of why I built the thing.

Both spec roles. As a host, /.well-known/ai-catalog.json on desvela.dev and desvela.ai, domain-anchored urn:air: identifiers, three entries, one of them the registry itself as application/ai-registry+json. As a registry, POST /search at registry.desvela.dev accepting query.text with filter.type, pageToken pagination and the federation modes. POST /explore (§7.3) is implemented as well, with facets over type, publisher and tags. Remote MCP lives at POST /mcp and exposes a free preflight(domain) tool.

Both sides pass the spec's own conformance CLI with 0 errors:

conformance-test manifest https://desvela.dev/.well-known/ai-catalog.json
conformance-test registry https://registry.desvela.dev

What it returns, and what it does not. Results are entries parsed from published manifests, carrying a relevance score and nothing that resembles a trust signal. I do not live-probe endpoints before indexing, so unlike Tiza a returned MCP server is declared rather than verified working. That seems worth stating plainly, given these results land in front of agents. What I do verify is that a surface is genuinely there: sites that answer 200 with HTML on any path are tested against a canary request and excluded, which removed roughly 16% of the apparent publishers.

Happy to adjust the description, or to be dropped from the list later if a neutral crawl turns out not to earn its place. That seems easier to find out with it in the menu than by guessing.