| name | tavily |
| description | Web search, content extraction, site mapping, crawling, and deep research via Tavily — use whenever the user needs current web information, wants to pull clean content from a URL, discover pages on a large site, bulk-extract documentation, or produce a cited multi-source research report. Prefer this skill for any agentic web-research workflow; do not fall back to ad-hoc browsing when Tavily can answer the question more cleanly. |
Tavily
Tavily turns the open web into LLM-ready evidence. Use it when plain search results are not enough: current info, domain-scoped discovery, clean page text, site structure, bulk docs, or cited synthesis.
This skill supports both interfaces Tavily ships:
- MCP server (
tavily-mcp) — tools like tavily-search, tavily-extract, tavily-map, tavily-crawl.
- CLI (
tvly) — same capabilities plus tvly research for deep multi-source reports.
Pick whichever your environment exposes. The decision flow and parameters are the same; only the call syntax differs.
Pick the right capability
Use the smallest capability that answers the question. Escalate only when needed.
Escalation path
Do not jump straight to crawl or research.
- search — when you do not yet know the right URL
- extract — when you know the exact page you need
- map — when one page is not enough and you need site structure
- crawl — when many pages from the same site are clearly required
- research — when you need a cited report synthesizing many sources
This keeps latency, cost, and context volume under control. Read the reference file for the step you land on; do not load all of them upfront.
Core guardrails
- Search first, then extract the best URLs. Do not search again for a page you already have.
- Prefer official domains for technical questions (
--include-domains or MCP equivalent).
- Keep result sets small unless breadth is explicitly part of the task.
- Use
--instructions + --chunks-per-source when crawling for agent context — returns relevant chunks instead of whole pages, preventing context explosion.
- Always cap crawls with
--limit and conservative --max-depth.
- Treat Tavily output as evidence to read, not a substitute for reading the source.
Defaults and noise
Tavily accepts defaults via DEFAULT_PARAMETERS (MCP) or CLI flags. Useful ones: max_results, search_depth, include_raw_content, include_images. Do not enable everything by default — larger responses mean more noise and more context spent.
See also