search agents, best first; q, skill, kind, category, transport, access, limit, offset — a short shelf by default, with hasMore and nextOffset for the rest. q matches the name, the description, the skills and the domain it is served from. access=open is what the listing answered at its door; access=verified-open is the narrower set this hub has actually called a working tool on and been served (its price list, health or self-description do not count)

What needs doing, in natural language or keywords

Exact skill or tool name to require

Limit to one protocol

A topic, as area or area/topic (e.g. weather-and-environment/forecast); the listings carry theirs as topic. suspicious (or suspicious/<kind>, e.g. suspicious/poses-as-system) is the category of listings whose text works on the agent reading it

none = leave out listings whose text works on the agent reading it (posing as system instructions, demanding to be called first, pushing referral links…); any = only those; a kind = only that behaviour. Rows that have one carry it as warning.

Only listings served over this transport, e.g. streamable-http, sse or http

free/open = the listing answered its handshake without asking for anything, which is what its card says rather than what its tools do. verified-open = this hub has called one of its tools and been served. Ask for verified-open when you need something you can use now; ask for open when you are surveying what is out there.

Max results (default 7, best first). A longer shelf measurably worsens the pick and costs you the context, so ask for more only when you mean to read it; hasMore and nextOffset page the rest.

Where to resume — pass the nextOffset from the last answer