/tavily-extract

name: tavily-extract
description: |
  Extract clean markdown or text from specific URLs via the Tavily CLI. Use this skill whenever a URL must be opened to be cited — papers, changelogs, release pages, blog posts, docs — including JavaScript-rendered pages. Processes up to 20 URLs per call and supports query-focused chunking for long pages.
allowed-tools: Bash(tvly *)

tavily extract

Extract clean markdown/text content from one or more URLs. In this repo, "opened" means extracted with this skill (or fetched with a built-in web tool) — only opened URLs are citable as evidence.

Before running any command

Check `tvly --status`. If `tvly` is not on PATH:

pip install -q tavily-cli || uv tool install tavily-cli

Auth comes from the `TAVILY_API_KEY` environment variable (already set in this environment). Never print it. If it is unset, report that and fall back to the built-in web tools.

Quick start

# Single URL
tvly extract "https://example.com/article" --json

# Batch (max 20 per call)
tvly extract "https://arxiv.org/abs/XXXX.XXXXX" "https://github.com/org/repo" --json

# Long pages: only the relevant chunks
tvly extract "https://example.com/docs" --query "release date changelog" --chunks-per-source 3 --json

# JS-heavy pages
tvly extract "https://app.example.com" --extract-depth advanced --json

Key options

| Option | Description | |--------|-------------| | `--query` + `--chunks-per-source` | Return only chunks relevant to a query (1–5 per URL) | | `--extract-depth` | `basic` (default, try first) or `advanced` (JS-rendered pages) | | `--format` | `markdown` (default) or `text` | | `--timeout` | 1–60 seconds for slow pages | | `--json` / `-o FILE` | Structured output / save to file |

Tips

• Always quote URLs (shell interprets `?` and `&`).
• Try `basic` first; escalate to `advanced` only if content is missing.
• arXiv abs pages often hide authors/dates in the extraction — cross-check exact

metadata with `curl -sL 'https://export.arxiv.org/api/query?id_list=...'` (see `radar-source-verify`).

• If a search already used `--include-raw-content`, skip the extract step.