Websearch MCP Server

An MCP server for web search and page fetching. Converts HTML to clean markdown using html-to-markdown (Rust-based, ~200 MB/s), cutting token usage by ~60% when LLMs consume web content.

No API keys required — search is powered by DuckDuckGo.

Token Savings

Run uv run python benchmark.py to reproduce. Results from fetching real pages:

At Sonnet pricing ($3/M input tokens), that's $0.72 saved per batch of 7 pages.

Tools

websearch_search

websearch_fetch_page

websearch_search_and_fetch

Setup

Requires uv.

uv sync

Usage

MCP Inspector (dev)

uv run mcp dev server.py

Claude Code

claude mcp add websearch -- uv run --directory /path/to/mcp-websearch-server python server.py

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "websearch": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/mcp-websearch-server", "python", "server.py"]
    }
  }
}

Benchmark

uv run python benchmark.py

Fetches a set of real pages, counts tokens with tiktoken (cl100k_base), and reports HTML vs markdown savings with cost estimates for Sonnet and Opus pricing.

Dependencies

• mcp — FastMCP framework

• httpx — async HTTP client

• html-to-markdown — Rust-based HTML-to-Markdown converter

• ddgs — DuckDuckGo search (no API key)

• truststore — system certificate store for SSL

• tiktoken — token counting (dev dependency, for benchmark)