The join is the product

None of these endpoints do anything a single upstream provider does on its own. Wikipedia doesn't know what X is saying about a claim. npm doesn't check the GitHub repo behind a package. A country's holiday calendar and its visa rules live on two different sites. That's the gap these fill: one call, several providers, one merged answer.

Ten ship today, grouped by the job they solve.

Verification and grounding

claim-corroborate takes a factual claim and runs it against web search, Wikipedia, X, arXiv, and a citation check at once, then rolls the result into a per-source stance (supports, contradicts, unclear) plus an overall corroboration score. Send it "the Great Wall of China is visible from space" and it comes back with the sources that debunk it and the snippet each one was judged on.

grounded-answer-check answers a question, or grades an answer you already have, against reranked evidence pulled from web search and Wikipedia. It returns a support score and the exact spans that had no backing evidence. Built for agents that need to know whether an answer is safe to repeat before they repeat it.

B2B identity

contact-golden-record takes 2 to 5 messy candidate records for the same contact and merges them into one deduplicated record with per-field provenance: which input each value came from, and a confidence score. Feed it two versions of "Bob Smith" with slightly different company names and get back one record plus a list of exactly which fields conflicted.

Trust and safety

x-account-authenticity scores whether an X handle is a real person or a bot before an agent follows it, replies to it, or trusts what it says. It pulls the profile and recent timeline, checks the text for AI-generated tells, and screens the handle and display name for brand-impersonation homoglyphs. It's a heuristic label, not a guarantee, and the response says so.

dependency-provenance checks a package before you install it: registry download stats, a risk score, and the health of the source repo behind it, auto-routed to npm, PyPI, or crates.io from the package name. Ask about hono and get back adoption numbers next to the GitHub repo's health grade and star trend.

Research

literature-landscape runs one query across arXiv, PubMed, USPTO patents, and grants.gov, dedupes the same work when it shows up in more than one corpus, and verifies the top hit. Built for prior-art searches and grant scouting where the same paper often turns up under three different IDs.

cross-lingual-topic-monitor watches a topic across non-English sources: it searches the web and X, detects what language each hit is actually in, translates it, and scores sentiment on the translated text. Ask about "lithium mining regulation" and get back a language mix breakdown alongside the translated hits.

Crypto data

token-attention-report pulls a ticker's headlines, social sentiment, Wikipedia interest trend, and realized volatility into one snapshot. It's strictly numbers and headlines, no price target, no buy or sell signal, and says so in the disclaimer field.

Ops

cross-border-ops-brief resolves a country and returns the FX rate, visa requirements for a given nationality, upcoming public holidays, and the current local time, all in one call. It's reference data, not money transmission or immigration advice.

property-diligence-pack takes a street address and returns geocoding, tax assessment, permit history, seismic activity, and air quality for the resolved coordinate. Tax and permit data only cover a handful of US open-data cities right now and degrade gracefully everywhere else; seismic and air quality work for any address on earth.

Why this shape works

Every one of these reuses endpoints already live in the catalog, so the marginal cost of shipping the composite is orchestration, not new data. Registry, cluster map, Bankr manifest, and MCP package output all carry all 10 slugs as of this update.