web-search-mcp

Why this exists

Claude Code routed through LiteLLM or AWS Bedrock has WebFetch (client-executed — reads a known URL) but not WebSearch (server-executed — discovers URLs). The distinction:

• WebFetch works on any provider because Claude Code itself makes the HTTP request.

• WebSearch is an Anthropic-run server-side tool that Bedrock doesn't support.

The only missing primitive is discovery — turning a question into a list of relevant URLs. This MCP server fills that gap with a single web_search tool backed by ddgs (DuckDuckGo metasearch). Claude Code's native WebFetch handles reading whatever pages it finds interesting.

Related MCP server: DuckDuckGo MCP Server

How it works

You ask Claude Code a question that needs live web info
    │
    ▼
Claude calls web_search("your query")
    │
    ▼
This MCP server queries DuckDuckGo via ddgs
    │
    ▼
Returns: titles + URLs + snippets
    │
    ▼
Claude uses WebFetch on the URLs it wants to read

Setup

Prerequisites

• Python 3.13+

• uv package manager

• Claude Code CLI

# Install uv (if you don't have it)
curl -LsSf https://astral.sh/uv/install.sh | sh

Install

git clone https://github.com/victormacaubas/web-search-mcp.git

Register with Claude Code

claude mcp add --scope user web-search -- uv run --directory /absolute/path/to/web-search-mcp python -m web_search_mcp

Then allow-list the tool in ~/.claude/settings.json:

{
  "permissions": {
    "allow": [
      "mcp__web-search__web_search"
    ]
  }
}

Restart Claude Code and the web_search tool is available in all sessions.

Usage

Once registered, Claude Code will automatically use it when it needs to search the web. You can also prompt it directly:

"Search for the latest Anthropic MCP documentation"

The tool returns JSON with results:

{
  "results": [
    {
      "title": "Model Context Protocol Documentation",
      "url": "https://modelcontextprotocol.io/docs",
      "snippet": "The Model Context Protocol (MCP) is an open protocol..."
    }
  ]
}

Parameters

Architecture

src/web_search_mcp/
├── __main__.py    # Entry point (python -m web_search_mcp)
├── server.py      # FastMCP instance + web_search tool
├── search.py      # SearchBackend protocol + DdgsSearchBackend
└── models.py      # SearchResult dataclass + WebSearchInput validation

Key design decisions:

• Swappable backend — SearchBackend is a Python Protocol. The ddgs implementation can be replaced with SearXNG or a licensed SERP API without touching the MCP layer.

• stdio transport — Runs as a subprocess of Claude Code. stdout is the MCP wire, stderr for logs.

• No fetch_url tool — Redundant with Claude Code's native WebFetch.

Development

# Run tests
uv run pytest

# Lint + format check
uv run ruff check . && uv run ruff format --check .

# Type check (strict)
uv run mypy src/

# Run the server directly
uv run python -m web_search_mcp

Swapping the search backend

The SearchBackend protocol has a single method:

class SearchBackend(Protocol):
    async def search(self, query: str, max_results: int, region: str | None) -> list[SearchResult]: ...

To use a different backend (e.g., SearXNG, Brave API), implement this protocol and swap the backend variable in server.py.

Disclaimer

This tool is designed for ad-hoc, conversational web searches — a developer occasionally checking docs, verifying a fact, or looking something up mid-session. It is not intended for:

• High-volume agentic workflows that issue dozens of searches per minute

• Production systems with uptime requirements

• Commercial applications at scale

The ddgs library scrapes DuckDuckGo's frontend. At conversational volume (a few searches per hour) this works reliably. At high volume, you will hit rate limits or CAPTCHAs. If your use case requires production-grade search at scale, swap in a licensed Search API via the SearchBackend protocol.

Limitations

• ddgs relies on scraping — If DuckDuckGo changes their frontend, searches break until the package is updated.

• Rate limits — DuckDuckGo can rate-limit or CAPTCHA under heavy load. At conversational volume this is unlikely.

• No guaranteed uptime — This is a personal tool, not a service.

License

MIT