Product
One index, reached two ways: a REST API and a first-party MCP server with the same filters and the same key. Everything below is on every plan, including the free one — tiers buy credits, archive depth and watch capacity rather than features.
Capabilities
The parts with a page of their own
Six of these are worth more than a row in a table, because each one is a specific failure of news retrieval rather than a feature name. The page says what goes wrong, what we do about it, and where the approach stops working.
Semantic search
Keyword, semantic and hybrid retrieval over the same index and the same filters.
A query has to guess the words a newsroom chose. Outside English it usually guesses wrong.
Story clustering
One event returns once, carrying the outlets that covered it.
One wire report arrives as forty results, and a model reads the repetition as corroboration.
Filters
Publisher, language, country, city, topic, entity and numeric ranges, with facet counts.
Narrowing by hand after the fact costs calls and still leaves the wrong set.
Signals
25 scored language dimensions across industry, business register and risk.
Subject and framing are different questions, and one keyword answers neither.
Relationships
Who is reported alongside whom, how two entities connect, and what a watchlist shares.
The interesting question is rarely about one name on its own.
Language coverage
Language as a first-class filter, and why script changes what retrieval can reach.
A filter that exists is not the same as retrieval that works in that language.
Reference
Everything the API does
The full surface, including the parts that live in the technical documentation rather than on a page here. Nothing in this table is gated behind a tier.
| Capability | What it is |
|---|---|
| Retrieval modes | Keyword, semantic and hybrid over one index and one set of filters. All three on every plan, including the free one. |
| Story grouping | A single event returns once, carrying the outlets that ran it, the languages they ran it in and when it started. |
| Exact filters | Publisher, language, country, city, topic, organisation, person, author and source adapter, with facet counts on the response. |
| Signal ranges | 25 scored language dimensions, filtered by numeric range and stacked with any other constraint. |
| Entity resolution | Resolve a name to the spellings the index holds, grouped with counts, then filter on the id rather than on a string. |
| Place filters | A radius around a point, or a bounding box. Each result lists the cities that matched, nearest first. |
| Event classes | 17 classes such as bankruptcy, outage and labour strike, derived from the topics an article's source supplies. |
| Relationships | Co-mentioned entities, connection paths, changing networks and shared exposures, each answer carrying its evidence articles. |
| Time series | Counts per hour, day, week or month — articles, distinct stories, distinct outlets and a signal’s percentiles — in one call. |
| Article and story lookup | Fetch one article by id, expand a story into every article behind it, or find the articles related to one. |
| Publisher profiles | Whether an outlet breaks stories or carries them, and how far behind it runs when it is not first, measured from its own publishing. Observations about behaviour, never a credibility score. |
| Provenance | Every article names the ingestion adapter and feed it came through, both filterable; every response carries the attribution its sources require. |
| Watches and webhooks | The index runs a story query on your interval and records new and growing stories as events, read as a feed or posted to a signed webhook. |
| Language coverage | Language is a filter rather than a flag on an English index, and retrieval behaviour is set per script. |
| MCP server | 18 tools over the same query surface as the REST API, authenticated with the same key. |
| Credits and previews | Every call is priced from one published table, and estimate=true returns what a call would cost without running it. |
Surfaces
Two surfaces, one implementation
The MCP server is not a wrapper around the REST API — the eighteen tools are generated from the same query surface the endpoints expose, so a filter that exists in one exists in the other and a query you debugged with curl behaves identically when a model issues it. One key works on both, one plan covers both, and usage is counted the same way.
That matters mostly when something goes wrong. A community wrapper around someone else's API is a second implementation of the same semantics, maintained by a third party, and the failures it introduces are invisible from either side of it.
The response
What we return
Headline, publisher, timestamp, language, link, a snippet, and a story identifier — not article bodies. Those stay with the publishers who wrote them, which keeps you clear of the licensing questions that follow full text around. If your application needs the text itself, license it from the publishers directly; this is the wrong foundation for it and no plan changes that.
Every response also says what it cost. A call is priced in credits from
one published table, and passing estimate=true returns the range a call would cost without running it, which is how you bound an
agent that reformulates its own queries.
Limits
Where this is the wrong tool
A news index tells you what is being reported. That is narrower than it sounds: it is silent on what is true, on what is planned, and on everything that happened without anyone writing about it. If your reader is a person scanning a feed and duplicate coverage does not bother them, the grouping this product is built around is solving a problem you do not have, and simpler options are cheaper. If you need newspaper front pages, or country filtering as the primary axis of your product today, the comparison pages name the vendors who are stronger there.