| name | tavily-overview |
| description | Entry point and routing guide for the Tavily web research toolkit. Loads when the user mentions Tavily, asks how to research the web, asks "what can Tavily do", or starts a research task without specifying which Tavily skill. Explains the five Tavily MCP tools, the workflow skills built on top of them, and routes to the right narrower skill for the user's task. Use this skill first when the right Tavily entry point is unclear. |
Tavily Overview
Tavily is a web research toolkit for Cowork. It exposes five MCP tools that
read the live web, plus workflow skills layered on top for common knowledge
work. This skill is the routing guide — read it when the user's intent is
clear ("research X", "compare A vs B", "what's happening with Y") but the
right Tavily entry point is not.
Connection check
Before invoking any Tavily skill, confirm the Tavily MCP is connected. The
MCP tool names look like tavily_search, tavily_extract, tavily_crawl,
tavily_map, and tavily_research (some Cowork installations namespace
them; trust the tool list, not the literal name).
If no Tavily tools are visible, tell the user: "The Tavily MCP isn't
connected yet. Open the plugin in Cowork → Connect → sign in to Tavily."
Don't fall back to other web tools — Tavily is the point.
The five raw MCP tools
| Tool | Use it for |
|---|
tavily_search | Find pages on a topic. Returns titles, URLs, snippets, and relevance scores. |
tavily_extract | Pull clean markdown from one or more known URLs. |
tavily_map | Discover URLs on a site without downloading content. Faster than crawl. |
tavily_crawl | Bulk-extract content from many pages on a site. |
tavily_research | Long-running deep-research agent that returns a cited multi-source report (30–120s). |
The skill map
For Cowork users you almost always want a workflow skill, not the raw tool.
Pick by what the user is trying to accomplish:
| User says… | Use this skill |
|---|
| "Search the web for…", "Find articles about…", "What's the latest on…" | web-search |
| "Read this URL", "Pull the content from…", "Extract this page" | extract-page |
| "What pages exist on…", "List URLs for…", "Find the page on this site about…" | site-map |
| "Download all the docs at…", "Bulk-extract everything under /blog/…" | site-crawl |
| "Research X", "Write a report on…", "Compare A vs B vs C" | deep-research |
| "Profile this competitor", "How does X price?", "Build me a battlecard" | competitive-intel |
| "Track news on…", "Daily digest of…", "What changed about X this week?" | news-monitor |
| "Is this true?", "Verify this claim", "Find sources that confirm/refute…" | fact-check |
| "What topics do competitors cover that we don't?", "Find content gaps" | content-gap-analysis |
| "Research this company before our call", "Brief me on this person" | lead-research |
| "Build a reading list on X", "Curate sources about Y" | source-curation |
Workflow patterns
Most research tasks fit one of these shapes. Default to the simplest one
that gets the job done.
- Single search: a quick fact, a recent headline, "who is X" →
web-search and stop.
- Search → extract: search for the right URL, then extract clean
content from the best result. Use when the user wants the content, not
just a snippet.
- Map → extract: the user knows the site but not the page. Map first,
then extract from one or two URLs.
- Crawl: the user wants every page in a section (docs site, blog,
pricing pages). One
tavily_crawl call beats dozens of extracts.
- Research: deep, multi-source, cited. Always slower (30–120s) but the
best answer for "compare X vs Y" or "give me a market report on Z."
Cost-and-time hints
tavily_search is fast and cheap. tavily_extract is fast. tavily_crawl
and tavily_research are slower and more expensive — surface that to the
user before kicking one off if they look like they're expecting an instant
answer.
Saving outputs
When the user asks for a deliverable (report, brief, digest), default to:
- Save the markdown to
~~document store if a document store is connected.
- Otherwise, write a
.md file to the workspace folder and link it.
- Always include a "Sources" section listing every URL referenced, with
citation numbers in the body.
See CONNECTORS.md at the plugin root for the full list of connector
categories the workflow skills reference.
