Find broadly
Search across models, harnesses, gateways, services, and runtimes without collapsing them into one ambiguous kind.
Find any model, harness, or AI-enabled service.
Give your agents a few precise commands to discover capabilities, inspect the evidence, and resolve a usable route.
The ambition is broad discovery. Every current result remains explicit about coverage, evidence, and unknowns.
Three commands between a name and a decision
Search across models, harnesses, gateways, services, and runtimes without collapsing them into one ambiguous kind.
Inspect the source, revision, digest, observation time, and verification state behind every useful claim.
Apply your written policy and receive one reproducible execution route — or a refusal that explains what is missing.
$ hiqs lookup "apex" --catalog fixture:synthetic
$ hiqs explain --claim c-apex1-release --catalog fixture:synthetic
$ hiqs resolve "model:fixture/apex-1@r1" --harness harness:fixture/forge-cli@1.4.0 --policy fixtures/strict-ranked.policy.json --catalog fixture:synthetic
→ one evidenced harness + gateway + model route, or a structured refusal
An identity, not an endpoint. Family, release, and how strongly it is pinned — artifact digest down to a floating alias, each with evidence for that classification.
$ hiqs lookup <model> --catalog fixture:synthetic
The thing that executes: an exact package revision, its execution mode, adapters, and the capabilities someone evidenced.
$ hiqs lookup <harness> --catalog fixture:synthetic
A serving path: operator, API revision, transports, and concrete bindings to the models it actually serves.
$ hiqs lookup <gateway> --catalog fixture:synthetic
An AI-enabled product that may invoke models you never chose — with operation-scoped disclosure of which ones, including honest unknowns.
$ hiqs lookup <service> --catalog fixture:synthetic
A local inference scaffold: the software that loads and serves models on your own machine or cluster, with its distribution and API dialects.
$ hiqs lookup <runtime> --catalog fixture:synthetic
Success means exactly one harness + gateway + model triple with its binding, adapter configuration, evidence references, and a deterministic trace. Anything less refuses, and says exactly why:
AMBIGUOUS_ALIASa colloquial name maps to two models your policy ranks equallyTIED_ROUTEStwo gateways serve the same model and no preference was declaredNO_SERVING_BINDINGthe model exists but this snapshot records nobody serving itUNVERIFIED_CLAIMthe route depends on claims nobody has verifiedCONFLICTING_CLAIMStwo records contradict each other; the registry never picks a winnerINSUFFICIENT_PINthe identity is a floating alias and your policy requires a stronger pinEvery decision-relevant field names a source, a recorder, and an observation time — or explicitly says it cannot.
Claims are scoped to the exact claim digest they support. Editing a value invalidates its verification.
Capture, verification, effective, and publication times are separate fields, always visible.
Unknown is a valid, displayable, refusable state. Guessing never is.
This is a beta. The catalog bytes are unsigned and not for production trust. Demo data is labeled Synthetic or Imported wherever it appears — fixture execution is never presented as verification of real vendors.
"Feasible per these claims" does not mean credentials, quota, live availability, or upstream behavior are guaranteed. The catalog reports facts and possibilities; your policy — written by a human, accepted explicitly — makes the choice.