omniseek_sources

List all sources — call this to ROUTE before searching.

Fully-qualified MCP name: mcp__omniseek__omniseek_sources (server name is omniseek; there is no omniseek-eye server).

BOUNDED ORIENT: a bare (no-arg) call does NOT dump every source's facets. It returns the routing VOCABULARY (available_domains / available_regions with counts) + the capabilities verb index + source_names (the bare inventory) + counts, so the orient payload stays small no matter how far the roster grows (brain_orient's lesson). The per-source FACETS (kind / domains / regions / modes, needs_credentials, explicit_only, stability, health, ...) plus the prose description arrive when you NARROW or ask verbose — reach for them on demand: • domain="jobs" / "papers" / … → only sources whose domains facet contains it, WITH their full descriptions. domain= is the most RELIABLE router; the no-arg call returns available_domains (the full closed vocabulary + counts) so you can pick a valid token, and a near-miss (e.g. "careers") returns did_you_mean instead of a silent empty. • query="singapore visa" → TOKEN-OVERLAP over name + description + domains + regions + cross-lingual keywords, ranked best-first (multi-word- and English↔中文-safe), WITH descriptions. • region="sg" / "ca" / "cn" → only sources whose regions facet contains it (the no-arg call returns available_regions; a near-miss returns did_you_mean). Region narrowing matters when the deployment's source pack is geographic. • verbose=True → the full unfiltered list, WITH every description. check_health=True does a fresh LIVE probe of every source (slow) AND returns a system block: the recall-index health (indexed_docs / embedder_available / vec_embed_failures / last_write_age_s) plus the observation-journal durability head, materialization cursor, pending count, and failures. and the openalex_usage attribution (which component spent the shared daily budget + remaining).

The no-arg (orient) call also returns capabilities: the non-search VERB index (field_skeleton, coauthors, transcribe, …) so you discover the whole toolkit here, not only after loading a tool.

kind per source: lookup (query = search the venue) | stream (query = FILTER over a recent feed; nowcoder additionally adds a site search) | proxy (search-index venue: engine snippet, shared paced backend) | portal (single-URL fetch). explicit_only_reason says WHY a source stays out of the broad sweep; a reason containing "shared paced backend" means the source draws on the one web-search backend of section (10) in the server instructions.

Returns: {"count": N, "backend_count": M, "backend_breakdown": {...}, and EITHER

• a BARE ORIENT: "source_names": [...] + "note" + available_domains + available_regions + capabilities; OR

• a NARROWED (domain/region/query) or verbose call: "sources": [{name, backend, (description when narrowed/verbose), needs_credentials, explicit_only, explicit_only_reason? (present only when excluded; the full catalog of why-strings search's _meta.excluded_count no longer re-ships), param_hint? (the structured query a VERTICAL source wants — a stock code / ticker / author name — present only when the source declares one, so a named call is filled right the first try), stability, access_tier, health, health_as_of, kind?, domains?, regions?, modes?, (healthy, status if check_health)}]. (did_you_mean on a domain/region near-miss; system:{recall, openalex_usage, jobs:[{name, schedule, enabled, last_run, next_run, budget_s, desc}, ...]} when check_health — the background-job fleet.)}

count is the RAW source count; it over-states coverage when many logical sources sit on ONE upstream. backend_count is the distinct UPSTREAMS (the honest figure) and backend_breakdown names every upstream backing >1 source, e.g. {"openalex": 42} (40+ affiliation slices of one corpus + one API budget + one breaker = one backend, not 40 of coverage).