Wikexa

你的AI智能体真正读得起的知识。每次你的智能体查阅资料，都要为约15,000个token的标记语言买单，却只换来大约500个token的事实内容。你为所有这些付费——导航、引文脚手架、编辑链接——然后你的模型把它们统统丢弃。Wikexa已经替你完成了那些阅读。你拿到的是事实。

GitHub

文档

Wikexa

Structured knowledge for AI agents. Wikipedia, Wikidata and Wiktionary as clean JSON instead of HTML. ~30x fewer tokens than fetching the page yourself.

Why

An agent fetching en.wikipedia.org/wiki/Apple_Inc. burns ~15,000 tokens of HTML to recover ~500 tokens of facts. Wikipedia's own API returns wikitext, which is barely better. One lookup() call here returns the facts directly.

Tools

ToolWhat it gives you
lookup(entity, corpus?)Structured facts + ~200-token summary. Resolves aliases and Q-ids.
article(title, sections?, max_chars?, corpus?)Full article as structured JSON — sections as an array, infobox as key/value. No HTML or wikitext.
define(word, language?, pos?)Dictionary definitions from Wiktionary — 8.15M entries across ~4,000 languages. Grouped by language, with etymology and pronunciation.
search(query, limit?, corpus?)Ranked matches blending relevance, pageviews, and exact-title matching. Up to 50 results.
recent(topic?, hours?, limit?)Recently changed articles — facts newer than your training cutoff.

Every tool is also a REST endpoint backed by the same handler:

GET https://wikexa.com/v1/lookup?entity=Tim%20Cook
GET https://wikexa.com/v1/article?title=Apple%20Inc.&sections=History
GET https://wikexa.com/v1/define?word=serendipity&language=English
GET https://wikexa.com/v1/search?query=quantum%20computing&limit=5
GET https://wikexa.com/v1/recent?topic=election

Example

curl -s 'https://wikexa.com/v1/lookup?entity=Tim%20Cook'

{
  "title": "Tim Cook",
  "wikidata_id": "Q265852",
  "description": "American business executive (born 1960)",
  "summary": "Timothy Donald Cook is an American business executive...",
  "infobox": { "born": "1960-11-01", "education": "Auburn University; Duke University", "title": "CEO of Apple Inc." },
  "facts": { "occupation": "business executive", "employer": "Apple Inc.", "citizenship": "United States" },
  "related": ["Apple Inc.", "Steve Jobs", "iPhone"],
  "source_url": "https://en.wikipedia.org/wiki/Tim_Cook",
  "license": "Wikipedia text CC BY-SA 4.0; Wikidata facts CC0"
}

Connect

Claude Code / any MCP client:

claude mcp add --transport http wikexa https://wikexa.com/mcp

Cursor / VS Code (mcp.json):

{
  "servers": {
    "wikexa": {
      "type": "http",
      "url": "https://wikexa.com/mcp"
    }
  }
}

Raw JSON-RPC:

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":"Apple Inc."}}}'

Discovery

PathWhat it serves
/.well-known/mcp.jsonMCP server metadata
/.well-known/agent.jsonAgent card
/openapi.jsonOpenAPI 3.1 spec
/llms.txtPlain-text orientation for LLMs
/.well-known/oauth-authorization-serverReturns 404 — no auth required

Coverage

  • 476,616 encyclopedia articles (top by pageviews)
  • 10.2M Wikidata entity records
  • 8.15M Wiktionary dictionary entries (~4,000 languages)
  • 194,870 records from Wikiquote, Wikibooks, Wikiversity, Wikivoyage
  • 1.9M full-text searchable with pageview-weighted ranking

Performance

Served from the Cloudflare edge. 5–12ms warm with zero storage operations. Verify with the x-wikexa-ops response header.

Pricing

Free for individuals and open-source projects. Commercial volume plans available at wikexa.com.

Token control

article() takes sections and max_chars. Ask for the section you need rather than the whole article. Empty fields are omitted from all responses to minimize token usage.

Licensing

Wikipedia text is CC BY-SA 4.0 — every response carries license and source_url. Wikidata facts are CC0. We add no restrictions on top of those licences. Paid plans buy service capacity and support, never content rights.

Wikexa is not affiliated with the Wikimedia Foundation.