fetchdown

Free web search + clean-markdown extraction for LLM agents. A self-hosted, no-API-key replacement for paid search/read services (Jina, etc.).

• Search the web via DuckDuckGo (ddgs)
• Read any URL — HTML (via trafilatura) or PDF (via pymupdf)
• Clean the output: nav, ads, citations, TOC dumps, and duplicate boilerplate stripped
• Fetches all results in parallel with a connection-pooled, browser-impersonating HTTP client (curl_cffi)

Three ways to use it: as an MCP server (for tool-calling agents), as a Python library (for code-executing agents), or as an HTTP service.

Install

uv sync                 # core (library + MCP server)
uv sync --extra http    # also installs Flask for the HTTP server

1. MCP server (recommended for agents)

Exposes two tools — web_search and read_url — over stdio.

uv run fetchdown-mcp

Register it with an MCP client. For Claude Desktop / Claude Code, add to your MCP config:

{
  "mcpServers": {
    "fetchdown": {
      "command": "uv",
      "args": ["--directory", "/Users/arshchawla21/Projects/Business/fetchdown", "run", "fetchdown-mcp"]
    }
  }
}

The agent then calls:

• web_search(query, n=5, fetch_content=True) — search + read top results in one call
• read_url(url) — read a single page you already have a link to

2. Python library

from fetchdown import search, fetch

# Search + fetch + clean, in parallel:
results = search("python asyncio tutorial", n=5)
for r in results:
    print(r["title"], r["url"], r["chars"], r["error"])
    print(r["markdown"])

# Just the hits, no content (fast):
hits = search("python asyncio", n=5, fetch_content=False)

# Read one page:
page = fetch("https://example.com/article")
print(page["markdown"])

Every result is a dict: {url, title, markdown, chars, elapsed, error} (or {url, title, snippet} when fetch_content=False). Failures never raise — markdown is None and error holds the reason.

3. HTTP service

uv run fetchdown-http          # dev server on :5000

• GET /search?q=<query>&n=5&fetch=1
• GET /read?url=<url>
• GET /health