Wikexa is a structured knowledge API for AI agents. It exists because the cheapest way to answer a factual question should not be to download a web page.
An agent asked "who runs Apple?" typically fetches en.wikipedia.org/wiki/Apple_Inc. That page is roughly 15,000 tokens of HTML: navigation, edit links, citation scaffolding, infobox markup, references. The answer inside it is about five words. The agent pays for all of it, on every call, and throws nearly all of it away.
Wikipedia's own API is better but still not built for this. It returns wikitext — templates, transclusions and markup that a model has to interpret before it can read a fact.
We parse the Wikimedia dumps once, ahead of time, and store the result as one compressed JSON object per entity. Templates are resolved, references stripped, infoboxes turned into key–value pairs, sections split into an array, and Wikidata claims mapped from P-codes to readable names. What an agent receives is the part it was going to extract anyway.
There is no origin server and no database. A Cloudflare Worker runs in the data centre nearest the caller and reads objects from R2, with every read passing through the edge cache. A warm entity costs one Worker request and zero storage operations, which is why the free tier can be genuinely free.
A tailer follows the Wikimedia EventStream and rebuilds articles within minutes of an edit. That is the one thing a language model cannot do for itself: know what changed after it was trained.
More languages, more of the sister projects, and better coverage of entities that have Wikidata facts but no English article. If you need something specific, say so — the roadmap is mostly driven by what people ask for.
Wikexa is built by the same people as cabrini.ai, a market-data API for agents. Same architecture, different audience. Wikexa is not affiliated with the Wikimedia Foundation, and we are grateful to the volunteers whose work makes it possible — if you find it useful, donate to Wikipedia.