Agentic Keychain

Assess two to five named capabilities side by side for the same inputs, with the same method and ak.offer.v1 offer shape as evaluate_capability. Use when: you already hold a shortlist of capability ids and want each one's decision, reasons and offer in one answer. Not for: finding candidates (search_capabilities) or a single capability (evaluate_capability). Input: the capability ids, plus the optional task, model, expected_runs, baseline_cost_estimate_usd, context and constraints that evaluate_capability reads. Output: ak.comparison.v1 with the results ranked by decision, lower-bound net saving, relevance and capability id, each with its reasons and an ak.offer.v1 offer.

Per measured model: expected saving per run, break-even runs and net saving over expected_runs, with intervals, for one capability and no task. Use when: you know the capability and want its economics on each model the registry measured. Not for: a use-or-not answer for a task (evaluate_capability) or several candidates (compare_capabilities). Input: a capability id; add model to keep one model, and expected_runs with baseline_cost_estimate_usd for an estimate at your scale. Output: ak.roi.v1 with one economics entry per measured model; break-even runs is null when the interval does not show a saving, and your estimate is left out where success is not accepted or your baseline lies outside the measured range.

Decide whether a registry capability is worth using or buying for your task, from registry-run benchmarks only. Use when: you have a task and want a recommend, do_not_recommend or undetermined answer with the arithmetic; name a capability to weigh only that one, or leave it out and the registry weighs its closest matches. Not for: one short answer without the arithmetic (resolve_task), a plain list of matches (search_capabilities), two to five named candidates side by side (compare_capabilities), or the break-even figures of one capability without a task (estimate_roi). Input: a plain task description; give model, expected_runs and baseline_cost_estimate_usd for a money estimate, and context for the conditions the capability declares. Output: ak.evaluation.v1 with decision, reasons, expected effect, price, confidence, operator_message, evaluation_id, next steps and an ak.offer.v1 offer (status, next_action, economics, the four evidence kinds listed apart). do_not_recommend is a normal answer.

Registry-run benchmark summaries (labelled), publisher claims (labelled PUBLISHER CLAIMED) and the signed attestations of one capability. Use when: you want the measurements behind an answer (paired runs, model, fixtures, method) or want to check them yourself. Not for: an answer for your task (evaluate_capability) or artifact hashes and the log proof (get_provenance). Input: a capability id, ak:// URI or slug. Output: ak.benchmark.v1 with the evidence, the publisher claims, the attestations, the method and the limitations; publisher claims never decide anything.

The capability's card: purpose, use-when and do-not-use-when conditions, permissions, price, evidence summary, trust checks, links and a task-independent ak.offer.v1 offer. Use when: you can name a capability and want its facts, or want to check the conditions an evaluation asked about. Not for: an answer for your task (evaluate_capability) or the full benchmark records (get_benchmark). Input: a capability id, ak:// URI or slug, from a search, an evaluation or a page. Output: ak.capability.v1 with the card fields, parameters, versions, terms, evidence, trust and links.

Artifact hash, manifest hash, signed publisher and registry attestations and a transparency-log inclusion proof for one capability. Use when: you want to check that a package is the one the registry published and signed. Not for: benchmark results (get_benchmark) or the card (get_capability). Input: a capability id, ak:// URI or slug. Output: ak.provenance.v1 with the hashes, the attestations, an RFC 9162 inclusion proof against the signed tree head, the trust checks and the keys URL.

A signed 15-minute price quote with the hosted checkout link. It charges nothing; a purchase requires operator approval in the hosted checkout. Use when: your operator wants the price and the checkout link of a capability that is for sale, for example after an evaluation answered recommend. Not for: deciding whether a capability is worth it (evaluate_capability); a free capability needs no quote. Input: the capability id and, if you evaluated first, the evaluation_id, which links the quote to that evaluation. Output: a quote_id and signed claims bound to the capability version, price, currency, terms and expiry, the checkout link and the unlock request; the price comes from the registry record only.

Report how a capability worked for you. Stored as AGENT REPORTED; it never changes evidence labels, trust or ranking. Only the listed fields are accepted; no task text is kept. Use when: you ran a capability, or decided not to, and can say whether the task succeeded. Not for: questions about a capability (get_capability, get_benchmark). Input: the capability id and success; add evaluation_id or purchase_ref to link the report, the optional token, tool-call, latency and cost counts, and failure_reason only with success false. Output: whether the report was stored, its label and, for a failed paid capability, the remedy policy.

Describe a task in plain words and get one compact answer: whether a registry capability is worth using for it, from registry-run benchmarks only, and the next step. Use when: you have a task and no capability name, and want one decision word (use, do_not_use, use_free_alternative, insufficient_evidence, blocked_by_constraints or no_match) with its reason codes. Not for: the full arithmetic, the offer and the operator message (evaluate_capability with the same inputs), a plain list of matches (search_capabilities) or a capability you can already name (get_capability). Input: task, plus optional model, expected_runs, baseline_cost_estimate_usd, context and constraints; any other field is refused with the allowed list. Output: ak.resolve.v1 with decision, capability id, reason codes, the measured saving per run with its interval or null, the estimate at your own baseline with its break-even runs or null, confidence, evidence label, price, free alternatives, next action, evaluation_id and the tool that gives the detail. do_not_use and no_match are normal answers.

Search capability names, summaries, tasks and use-when conditions by keywords, ranked by lexical relevance only, never by payment. Use when: you want to see which capabilities exist for a topic, or you know part of a name. Not for: deciding whether one fits your task (resolve_task, or evaluate_capability for the full arithmetic) or reading a capability you can already name (get_capability). Input: search words, at most 500 characters; no earlier call is needed. Output: ak.search.v1 with id, name, summary, relevance, state, price and the strongest evidence label of each result, and no decision.

Exchange the license key from the checkout (a purchase requires operator approval) for a signed entitlement and a download link that lasts at most 15 minutes. How the key is handled: the privacy page at /legal/privacy/. Repeating it with the same key is safe and never charges. Use when: your operator completed the hosted checkout and gave you the license key. Not for: a price or a checkout link (get_quote); a free capability downloads from its page. Input: the capability id and the license key, plus quote_id and evaluation_id when you have them. Output: a signed entitlement, a purchase_ref for report_outcome and the download link.