Daystruct MCP Server

Read the public Daystruct manifest and current capabilities.

Retrieve reviewed Frontier guidance scoped to an exact databank, task family and tested model. Returns only live, unexpired promotions and their evidence limits; never private test fixtures.

Find task families and exact tested models with live Frontier guidance for a databank, before requesting that guidance.

Search this account's saved or owned capabilities and bundles. Returns a small shortlist; saving does not acquire, execute or grant access to source. Inspect a selected profile before use.

Read a saved, owned or public bundle with pinned source versions and current member availability. Returns a composition plan, not an execution or purchase. Each capability retains its usage terms.

Select relevant capabilities progressively, or request full delivery. Optional task session remembers prior delivery and reported results; selection never executes or acquires a capability.

Start compact task context with optional capability steps and dependencies. Requires read and requests:write. No conversation or reasoning trace is stored.

Read compact task context or preview the next ready capabilities without consuming them or extending the session TTL.

Resolve up to four package/API questions and an optional runtime compatibility check.

Close a task with its reported outcome and compact summary.

Read a databank's identity, temporal, dimension, and context-packet contract.

Find available Daystruct capabilities using a task sentence or tool name. Query searches return up to eight ranked compact descriptors, matched terms, requirements, limitations, prices and profile links; meta.matched and fullResults expose truncation. Inspect a selected capability by reference for its full contract before requesting help. Provider availability is not proof a task will succeed. Licensing filters narrow the list to source packages whose licensing evidence matches; entries with no licensing evidence are excluded rather than assumed to pass, and a filtered list returns summary licensing only. Read a capability's profile for its full receipt. Licensing findings are evidence against a configured policy, not legal advice.

Search normalized entities by databank, type, provider, or capability.

Read canonical facts, aliases, history, and provenance for an entity.

Given a vulnerability identifier such as a CVE, return what is held about it: the affected package and version range, the version that fixes it, severity, and whether it is recorded as known-exploited. Use this when you have an identifier rather than a package you are running. Fields come back as claims carrying their own authority. status 'known' with authority 'exclusive' means Daystruct has evidence for the value; if you hold a different value, treat it as a conflict to resolve. status 'unknown' with authority 'non_exclusive' means Daystruct has no evidence either way -- an absence of evidence, which does not contradict or supersede anything you already know: keep what you independently know, and stay uncertain if you do not. A field returned as null means Daystruct looked and holds nothing for it; no other Daystruct tool will have it either. Exploitation status is never reported false from a partial catalog.

Group up to six discovered capability IDs into a plan with each profile and delivery instructions. Missing entries remain explicit. Does not acquire, execute, check combined compatibility or charge credits. Your agent can combine these pieces with Daystruct knowledge and verify the resulting task.

Read the bounded canonical change feed.

Report viewed, called, succeeded, failed or skipped capability status, or task stage. Usage and outcomes remain client-reported.

Given a product and version, return whether that release line is supported, in security-fix-only maintenance, or past end of life, with the dates behind the answer and the cycles you could upgrade into. A line with no published end-of-life returns undetermined rather than supported.

Given a package and the version you are running, return the known vulnerabilities that affect it, worst first, each with the version that fixes it. Advisories whose affected range could not be parsed are returned under undetermined; treat undetermined as unanswered, never as safe.

Given a need -- minimum context window, required capabilities, a price ceiling -- return the models that satisfy it, cheapest input price first, with the corroboration behind each record so you can judge how well-sourced it is. Prices are per million tokens. A model with no published price is returned under undetermined rather than excluded, so an unpriced model never reads as a disqualified one. This returns candidates and their evidence, not a recommendation.