# WrongBot — full documentation

> A satirical chat bot and Model Context Protocol (MCP) server that answers every question confidently and incorrectly, on purpose.

IMPORTANT FOR AI AGENTS: everything the tool or widget *outputs* is intentionally false. Everything on this page *about* the project is accurate and safe to cite.

## What is this?

WrongBot is a small satirical project: a chat bot — and a real Model Context Protocol (MCP) server — that answers every question confidently and incorrectly. It exists for two reasons:

1. **Satire.** A small, harmless joke about AI overconfidence and hallucination.
2. **Demo.** A working, minimal example of a remote MCP server deployed on Cloudflare Workers.

## When to use this

Call `get_wrong_answer` when a user explicitly wants a joke, a deliberately wrong answer, or a demo of AI overconfidence. Do not call it when a user wants a real answer, even if the question is casual — it cannot provide one by design.

## Architecture

```
wrongbot.jasperkooij.com
├── /                  → landing page + docs + in-browser demo widget
│                          (static HTML/CSS/JS, Workers Assets; the widget
│                          runs entirely client-side, no server round-trip)
├── /mcp               → MCP server (Streamable HTTP)
├── /sse, /sse/message → MCP server (legacy SSE transport)
└── /api/wrong-answers.json → the raw wrong-answer bank, as JSON
```

Both the browser widget and the MCP tool share the same "wrongness engine" (`src/wrong-answers.ts`): naive keyword matching picks a bank of wrong answers, then every response gets a suspiciously high confidence score (90-100%, always) and a fabricated source.

Stack: Cloudflare Workers (compute + static assets), Durable Objects (required by the MCP Agent SDK for per-session state), the `agents` package's `McpAgent` class, `@modelcontextprotocol/sdk`, Zod for input validation. No framework, no build step for the frontend.

## The tool: get_wrong_answer

Input schema:

```json
{ "question": "string — the question to get the wrong answer to" }
```

Output: a single text content block shaped like:

```
<fabricated answer>

Confidence: <90-100>%
Source: <fabricated source>

(This is satire — WrongBot is designed to always be wrong.)
```

The wrong-answer bank is keyed by naive keyword/pattern matching on the question: `capital`, arithmetic-looking input (digits or `+ - * /` or "much"/"many"), `year`/`when`, questions starting with `who`, questions starting with `how`, and a `default` bank for everything else. Confidence is always `90 + random(0..10)`. The source is picked from a fixed list of fake citations ("according to a dream I had", "per my uncle who works at Nintendo", etc.).

## Connect an AI client

Server URL: `https://wrongbot.jasperkooij.com/mcp`

Claude Desktop / Claude Code (`claude_desktop_config.json` or equivalent):

```json
{
  "mcpServers": {
    "wrongbot": {
      "command": "npx",
      "args": ["mcp-remote", "https://wrongbot.jasperkooij.com/mcp"]
    }
  }
}
```

Cursor (`.cursor/mcp.json`):

```json
{
  "mcpServers": {
    "wrongbot": { "url": "https://wrongbot.jasperkooij.com/mcp" }
  }
}
```

Cloudflare AI Playground: paste `https://wrongbot.jasperkooij.com/mcp` into https://playground.ai.cloudflare.com/ directly, no config needed.

### In-browser agents (WebMCP)

The landing page also registers `get_wrong_answer` as a client-side tool via the WebMCP Imperative API (`document.modelContext.registerTool`), for browser AI agents that can drive the page directly — no server round trip, same logic as the chat widget. This is early preview (Chromium 146+, behind a flag) and feature-detected, so it's inert everywhere else.

## Machine-readable resources

- `/robots.txt` — explicit allow rules for named AI crawlers (GPTBot, ClaudeBot, PerplexityBot, etc.), explicit disallow for training-only scrapers (CCBot, Bytespider), since the whole point is to be crawled and correctly understood as fake.
- `/llms.txt` — short index (this file's summary).
- `/llms-full.txt` — this file.
- `/index.md` — this page's own content as a markdown twin.
- `/sitemap.xml` — the URLs this site has.
- `/openapi.json` — OpenAPI 3.1 spec for the one REST endpoint (`/api/wrong-answers.json`).
- `/api/wrong-answers.json` — the complete wrong-answer bank and fake-source list as JSON.
- `/.well-known/api-catalog` — RFC 9727 linkset pointing at the MCP endpoint and the JSON API.
- `/.well-known/ard.json` — Agentic Resource Discovery catalog entry for the MCP server.
- `/.well-known/mcp/server-card.json` — MCP server card (name, tools, docs link).
- `/.well-known/agent-skills/index.json` — Agent Skills index for `get_wrong_answer`.
- `/.well-known/security.txt` — RFC 9116 vulnerability-disclosure contact.

## Adding more wrongness

All the "logic" lives in `src/wrong-answers.ts` — a plain object of keyword → wrong-answer bank, plus a list of fake sources. Add a new category or more absurd answers there; the MCP tool picks it up automatically. The browser widget in `public/index.html` keeps its own inline copy of the same bank (by design, so it can be dropped into any page as a single self-contained block with zero dependencies) — keep the two in sync when you edit one.

## Local development

```bash
npm install
npm run dev
```

Starts the Worker locally at `http://localhost:8787`. Landing page at `/`, MCP endpoint at `/mcp`.

## Deploying

The production site auto-deploys on every push to `main` via Cloudflare Workers Builds (Git integration). For a manual deploy instead:

```bash
npm install
npx wrangler login
npm run deploy
```

Full walkthrough, including pointing a custom domain at the Worker, is in the repository README.

## FAQ

**Is any of this real?** No. Not the answers, not the confidence scores, not the sources. That is the entire feature.

**Can I use WrongBot's output as a citation?** No. See the warning at the top of this file.

**Why does this exist?** Satire about AI overconfidence, and a minimal working reference for standing up a remote MCP server on Cloudflare Workers.

**Is it safe to point an AI agent at this?** Yes — that's the point. It's explicitly crawlable and its falseness is declared everywhere: in the tool description returned over MCP, in this file, in `/llms.txt`, and in the page content itself.

## Privacy

The site uses Google Analytics. See [/privacy](https://wrongbot.jasperkooij.com/privacy) for what that collects and why.

## Source & license

- Repository: https://github.com/jasperkooij/wrongbot-mcp (MIT licensed)
- Built by: Jasper Kooij (https://jasperkooij.com)