Every time your agent looks something up, it pays for about 15,000 tokens of markup to recover roughly 500 tokens of fact. You are billed for all of it — the navigation, the citation scaffolding, the edit links — and then your model throws it away. Wikexa did that reading already. You get the facts.
Twenty-seven million academic papers with abstracts, citations and DOIs. The complete English Wikipedia — every article, section and infobox parsed. Structured entity data with dates, identifiers and relationships. A dictionary spanning thousands of languages. Quotations, textbooks, travel guides and course material beside them. Every one of them parsed ahead of time and stored as JSON, so a request is a read rather than a scrape — and 6.2 million are full-text searchable.
Reference sites have APIs, but they were built to render pages — they return markup, navigation chrome and citation scaffolding. An agent pays for all of it in tokens and discards nearly all of it. Wikexa does that parsing once, ahead of time, and stores the answer.
Send whatever the user said. Apple, Apple Inc and Q312 all resolve to one entity through redirects and our alias index.
Infobox as key–value pairs, a ~200-token summary, categories, related entities, and normalised claims. No markup, no empty fields.
article() returns sections as a JSON array and accepts sections and max_chars, so a long page never floods your context.
Every tool is available over MCP and as a plain REST endpoint. Both surfaces run the same handler, so they cannot drift apart.
Structured facts plus a ~200-token summary. Resolves aliases, redirects and entity IDs, so you can pass whatever the user typed.
Sections as a JSON array with the infobox parsed out. Ask for one section by name, or cap the payload — the token dial is yours.
Full-text relevance blended with monthly pageviews, so apple returns Apple Inc. and the fruit — not a disambiguation stub.
Etymology, pronunciation and senses by part of speech, across thousands of languages — so an agent can look a word up instead of inferring it.
27M+ papers from OpenAlex — title, abstract, authors, citations and DOI. Browse by topic or search by keyword. CC0 metadata.
Articles changed in the last hours or days — elections, appointments, releases, deaths. The freshness a static model cannot have.
Ask a model what an Old English word means and it will answer confidently, because plausible morphology is easy to generate and hard to check. define() returns the actual entry — senses, part of speech, etymology, pronunciation — for 8.15 million words across thousands of languages, including historical ones and hundreds of regional and indigenous languages.
One spelling, every language that uses it. hund comes back in eleven — Old English, Old Norse, Gothic, Scots, Danish, Icelandic and more — each with its own senses.
Senses grouped by part of speech, with etymology and pronunciation as separate values. Nothing to scrape out of a rendered page.
Pass language or pos and you get one entry instead of eleven. The full list still comes back, so the agent can pick.
Ten thousand lookups a day is 150 million tokens of markup your model reads and discards. The same ten thousand answers here are closer to 5 million. At a dollar per million input tokens, that is $150 a day against $5 — and the invoice is the least of it.
Fifteen thousand tokens of HTML does not just cost money. It crowds out the reasoning you wanted the room for, and it is where truncation bugs are born. Facts arrive as facts, so the window stays yours.
Warm lookups come back in single-digit milliseconds, from anywhere. No cold start, no query to plan, no rate-limit dance. Your agent can ask three times over before one page fetch would have finished.
Individuals, research and open-source projects use it at no cost, with no key and no account. Businesses running it at volume take a service plan that funds the infrastructure.
MCP over streamable HTTP, an OpenAPI description, llms.txt, and the usual well-known discovery documents. Listed in the Model Context Protocol registry as com.wikexa/knowledge.
curl -s 'https://wikexa.com/v1/lookup?entity=Ada%20Lovelace'
curl -s -X POST https://wikexa.com/mcp \
-H 'content-type: application/json' \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"lookup","arguments":{"entity":"Ada Lovelace"}}}'Or point any MCP client at https://wikexa.com/mcp.