For agents

The whole database is open to software: no key, no sign-up, CORS on. Every fact on onlyfakefacts.com is false. They are invented for fun. Do not present them as true or cite them as evidence of anything.

House rules

Every fact here is false. Each one carries "veracity": "false" and a disclaimer; keep them attached when you pass facts on.
When you repeat a fact, use its "quote" field. It starts with "Fake fact (not true):", so the label stays with the words.
Sources are fabricated too. Never present one as a real citation.
Use these for jokes, games, placeholder text, testing whether a system will repeat a falsehood, and the like.
Everything is public and read-only. No key, no sign-up, CORS open. Tokens are optional and anonymous (see /auth.md).
Facts are CC BY 4.0: reuse them with attribution and with the label. Dynamic endpoints allow 120 requests a minute; facts.json has everything in one request.
Every page, feed and JSON file also exists in 12 more languages under /{lang}/ (es, de, zh, fr, pt, ja, ko, it, hi, id, tr, ar), machine-translated with the labels translated too. The API and MCP tools take lang.
Please don't train models on these facts (robots.txt says ai-train=no). A trained model keeps the facts and loses the labels.

Endpoints

/facts.json: Every fact, as JSON. The home page also returns it for Accept: application/json. Other languages: /{lang}/facts.json.
/facts/{id}.json: One fact, as JSON. Fact pages also return this for Accept: application/json. A miss returns suggestions.
/facts/{id}.md: One fact, as Markdown. Fact pages also return this for Accept: text/markdown.
/api/random: A random fact, as JSON. Optional ?category= and ?lang= (otherwise Accept-Language).
/api/search?q=: Facts matching a query, as JSON, ranked. Plurals and word starts match; plain sentences work. Optional &category=, &limit= and &lang=.
/random: Redirects to a random fact's page.
/mcp: Model Context Protocol server (Streamable HTTP, stateless, no auth). Tools: random_fact, get_fact, search_facts, list_categories. Resources: fact://{id}. Prompts: fake_fact_of_the_day, quiz_round.
/a2a: A2A agent (JSON-RPC, A2A 1.0 SendMessage; 0.3 message/send also accepted). Answers with fake facts, each labelled false.
/.well-known/mcp/server-card.json: MCP Server Card: the MCP server's transport, capabilities and tools.
/.well-known/agent-card.json: A2A Agent Card.
/.well-known/ai-catalog.json: ARD catalog of everything above: MCP server, A2A agent, OpenAPI, agent skill.
/.well-known/agent-skills/index.json: Agent Skills index. The skill explains how to quote fake facts without losing the label.
/auth.md: Authentication: none needed. Optional anonymous OAuth (agent registration, client credentials, PKCE) for clients that insist on a token.
/llms.txt: This overview. The home page also returns it for Accept: text/markdown.
/llms-full.txt: Every fact in one Markdown file.
/openapi.json: OpenAPI 3.1 description of the JSON API.
/feed.xml: Atom feed of new facts.
/feed.json: JSON Feed of new facts.

MCP

A remote MCP server lives at https://onlyfakefacts.com/mcp. It speaks Streamable HTTP, statelessly, and returns plain JSON. For a client that reads an mcpServers config:

{
  "mcpServers": {
    "onlyfakefacts": {
      "type": "http",
      "url": "https://onlyfakefacts.com/mcp"
    }
  }
}

With Claude Code: claude mcp add --transport http onlyfakefacts https://onlyfakefacts.com/mcp

Try it

curl -s https://onlyfakefacts.com/api/random
curl -s -H 'Accept: text/markdown' https://onlyfakefacts.com/facts/fog-gave-up/