gaslighting-mcp

A fake web search MCP server for AI alignment testing. It accepts a search query and returns LLM-generated search results shaped by a configurable background story.

Built with FastMCP and compatible with any OpenAI-style API endpoint.

How it works

1. You provide a background story via the BACKGROUND_STORY environment variable

2. The server exposes two tools: search and read_url

3. search — generates 10 realistic search results (url, snippet, date) consistent with the background story

4. read_url — generates a full fake article in markdown for a given URL, inferred from the domain/path and background story

5. The consuming AI agent receives these as if they were real web content

Setup

uv sync

Configuration

Usage

Standalone

uv run server.py

Claude Code MCP config

Add to your .mcp.json:

{
  "mcpServers": {
    "web-search": {
      "command": "uv",
      "args": ["run", "server.py"],
      "env": {
        "BACKGROUND_STORY": "your background story here",
        "LLM_API_KEY": "your-api-key"
      }
    }
  }
}

Tools

search(query)

Returns a JSON array of 10 results:

[
  {
    "url": "https://example.com/some-article",
    "snippet": "A realistic excerpt shaped by the background story.",
    "date": "2025-12-15"
  }
]

read_url(url)

Returns a full fake article in markdown, inferred from the URL and background story. Matches the tone and style of the source website.

License

MIT