An AI that answers questions about the world needs to know what happened today.
Unzoi turns the world’s news into something software can use: each story once per language, with who reported it, when, how widely and in which languages, delivered straight into the AI tools people already work in.
Why now
Three things changed at once
AI is being asked about today
A model knows the world up to the day its training ended. Assistants, copilots and agents are now expected to answer and act on what happened this morning, and they get it wrong when they guess.
Web search was built for people
A results page assumes a person will skim twenty links and notice that most are the same story. Software reads all twenty, pays for every one, and has no way to tell who reported first or whether anyone confirmed it.
Publishers are drawing lines
How AI companies use news text is contested. Unzoi returns facts about the reporting, short snippets and links, not the articles themselves. That is a deliberate boundary; it does not on its own settle every use, and customers still need terms that fit theirs.
Who it is for first
Teams shipping AI that gets asked about current events
The first buyer is the engineer behind a research copilot, an analyst assistant or an agent that has to answer about this week: someone already paying for a web search API and still getting stale or duplicated answers. They can try it in minutes, pay by use, and grow into it as their product does.
Monitoring, risk and media teams use the same index today through their engineers. Governments and regulated industries want it inside their own infrastructure, which is planned and not yet started. Both build on the same core and ask more of it: alerts that fit an analyst’s workflow, and deployment the customer controls.
What we are building
The layer between the news and the AI that needs it
Collecting articles is the easy part, and it is not where the value is. The value is in what sits on top: recognising that two hundred articles are one event, knowing which outlet had it first and how far it spread, knowing which companies, people and places it involves and how they connect, and knowing how each publisher behaves over time.
That is what Unzoi builds, and it reaches customers in the form AI tools now expect: an official connector for Claude, ChatGPT, Cursor and any app that supports MCP, plus a web API for everything else, on one key and one price list.
Today it runs as a hosted service. It is designed to run where a customer needs it to as well, inside their own cloud or a national network, which matters to governments and regulated industries that will not send their questions to someone else’s servers; those deployments are planned, and marked so below.
See it rather than read about it
The product, running now
The live demo answers from the production service while you watch, and a real AI agent is run against it every few hours with the transcript published as it happened.
What compounds
What gets better the longer it runs
- The record of what was reported (live). Every story, the outlets that ran it and when, kept and queryable over time with our grouping and entity links attached, rather than raw articles someone has to process again.
- How publishers behave (being built). Which outlets tend to be first and which follow, measured from their own publishing; it needs months of history to mean anything.
- What agents actually pick (live). The API records which result an agent or user chose, so ranking can be fitted to real use. The recording is live; fitted ranking does not exist yet.
- Where agents already are (live). An official connector in the AI tools developers use, priced per call, so adoption does not wait for a sales cycle.
These are reasons to expect an advantage, not measurements of one. The measurements we will publish are listed on the evaluation page.
Where it stands
What is live, and what is not yet
The same status list the documentation uses, last checked against the engineering roadmap on 20 September 2026. A roadmap with nothing marked “planned” would be the less believable one.
| Part | What it does | Status |
|---|---|---|
| Stories | One event returns once per language, carrying every outlet in the index that covered it. Clustering is ours, and it runs on articles from any source. | Live |
| Entities | A name resolves to the spellings the index holds, grouped under one id you can filter on. | Live |
| Relationships | Who is reported alongside whom, how two entities connect, and what a watchlist shares. Every answer carries its evidence articles. | Live |
| Signals | Theme scores per article, such as energy, litigation or supply disruption, measured as matching terms per hundred words and filterable by range. | Live |
| Watches | A story query the index runs on your interval, with an event feed and signed webhooks. | Live |
| Provenance | Every article names the adapter and feed it came through. Every response carries the attribution its sources require. | Being built |
| Native enrichment | Entities, places, dates, amounts and signals computed by us from the article text, so a record from any source carries the same fields. | Planned |
| Publisher profiles | Whether an outlet tends to be first in our index or follows others, how far behind it runs, what it covers and in which languages: measured from its own publishing, not asserted. | Being built |
| Cross-lingual event identity | One event, one story, across languages — so regional and national coverage of the same thing arrive together. | Planned |
| Event timelines | First report, which copies were syndicated from which original, and how coverage spread between sources over time — so an outlet list becomes a propagation path rather than a count. | Planned |
| Change detection | Watches that fire on a first report or a jump between countries, not only on a count that moved. | Planned |
| Claims | Who said what about whom, recorded as a statement with the articles that carry it, so a question about where a claim came from has a structural answer rather than a co-mention. | Planned |
| Unzoi Cloud | The hosted API and MCP server. What every plan on the pricing page runs on. | Live |
| Your cloud | The same build in an account you own, reading and writing storage you control. | Planned |
| Private network | No public endpoint: the API reachable only from inside your network. | Planned |
| Disconnected | Ingest from sources inside your perimeter, with no egress at all. | Planned |
Live: In production, on the public API, documented. Being built: Merged or in progress, not yet true of the hosted index. Planned: Designed far enough to state what would have to be measured. Not started.
What we will not claim
Reported is not the same as true
Unzoi tells you what was reported, by whom and when. It does not score whether a story is true or whether an outlet is reliable, and it will not pretend to. The measurements of quality we intend to publish, and what each will count, are listed on the evaluation page before any of them exist.