webpage-readability-mcp

Read-only MCP server that fetches a single URL and extracts its main content -- the article/post text, stripped of navigation, ads, sidebars, and boilerplate -- using trafilatura. This is the "readability mode" pattern (Firefox Reader View, Safari Reader, the original Arc90 Readability bookmarklet) as an MCP tool.

Tools

Setup

cd webpage-readability-mcp
python -m venv .venv
.venv/Scripts/activate   # Windows; use `source .venv/bin/activate` on macOS/Linux
pip install -r requirements-dev.txt

Run the tests

python -m pytest -q

Register with Claude Code

claude mcp add webpage-readability -- E:/repo/unt/mcp/webpage-readability-mcp/.venv/Scripts/python.exe E:/repo/unt/mcp/webpage-readability-mcp/server.py

Or add it manually to your MCP config (e.g. .claude/settings.json or the global Claude Code MCP config):

{
  "mcpServers": {
    "webpage-readability": {
      "command": "E:/repo/unt/mcp/webpage-readability-mcp/.venv/Scripts/python.exe",
      "args": ["E:/repo/unt/mcp/webpage-readability-mcp/server.py"]
    }
  }
}

Limitations

• No JavaScript rendering -- this is a plain HTTP GET + HTML parse, not a headless browser. A JS-heavy SPA whose content is injected client-side will extract poorly or empty.
• No authentication, cookies, or paywall handling -- every fetch is anonymous; a paywalled or login-gated page returns whatever the anonymous response contains.
• Single-fetch-per-call only -- exactly one HTTP request per call, to exactly the URL given. No pagination-following, no link-following, no caching.
• No PDF or other non-HTML content extraction -- a URL that doesn't resolve to an HTML/XHTML content-type returns an error rather than being silently mis-parsed.