tavily-search

Tavily search

Use the tavily_search, tavily_extract, tavily_map and tavily_crawl tools if your tool list has them (some clients add a prefix such as mcp__tavily__). Otherwise run tvly <command> directly; there's no need to check for it first. If neither works, follow setup.

Parameters are named as in the API (max_results). The CLI spells them as flags (--max-results) with comma-separated lists (--include-domains sec.gov,reuters.com). For exact parameters, read the tool schema, run tvly <command> --help, or see the API reference.

Search

Use advanced search. Set search_depth to advanced (--depth advanced in the CLI) for every search unless the user needs a fast answer to a simple lookup. It reaches more sources and ranks the passages that match your query.

Write queries for a search engine. Use the names and terms you expect on the page, not the user's whole request. Split a question with several parts into one focused search per part. For a long written report, search each part separately and extract the main sources for each before you write.

Follow your leads. When a result names a likely answer, test it right away: search that name with the other details you know, or extract its page. Keep a lead that fits most details until a source contradicts it; a detail you can't confirm doesn't rule it out. When a search misses, search the details again in their own words before adding names or filters you guessed, and build lists of candidates from results, not memory. When a detail keeps coming back empty, work out what it refers to, such as the role, event or term, and search for that.

tvly search prints each result's title, full URL and the first 600 characters of its matching passages. Add -o .tavily/<name>.json to also save every result whole, so you can go back to it instead of searching again. --max-chars 0 prints the passages whole. Read a promising page in full with tvly extract <url>.

Batch independent searches. Issue the searches for one step together in one turn, one call each: separate tavily_search calls with MCP tools, or separate Bash calls with the CLI, each holding a single tvly command. MCP calls run in parallel; Bash calls run one after another but still use one turn. Avoid &, $(...) and loops, which make the client ask for permission. Put the query in single quotes: tvly search 'Rust 2024 edition release date' --depth advanced --max-results 10 -o .tavily/rust-2024.json.

Work from saved files. Keep results in .tavily/ in the current folder and read them with head, grep or python3. -o creates the folder and keeps it out of git, so don't create it yourself. For example, to find a name or number in a long page:

tvly extract 'https://example.com/report' -o .tavily/report.json
python3 -c 'import json,re; t=json.load(open(".tavily/report.json"))["results"][0]["raw_content"]; [print(t[max(0,m.start()-200):m.end()+200], "\n---") for m in re.finditer("2019", t)]'

Read pages with tvly extract, not curl or wget.

Ask for 10 results. Set max_results to 10 (--max-results 10). Narrow before you widen. Filters help more than raising max_results further. Use time_range or dates for recent events or a stated period. Use include_domains when the answer lives on an official or named site; it is a hard filter, so search again without it if it doesn't find the page.

Keep large results on disk. When a result is long, such as raw page content or a crawl, save it under .tavily/ and read only the parts you need. An MCP result lands straight in your context, so keep those calls small.

Save copies with -o. Saving results to files is a job for tvly: run the call with -o .tavily/<name>.json, even if you searched with MCP tools. A file you write from an MCP result is your retyping of it, not the original. Without tvly, work from the results in your context and don't copy them into files. If the user asked for a saved copy, give them the URLs and queries, say an exact copy needs the CLI, and write a retyped copy only if they ask for one.

Read the results critically. Count independent sources, not results. Treat a generated answer as a summary to check, never as a source.

Extract

Extract returns up to 20 URLs per call as markdown, including pages built with JavaScript.

• Extract only the few results worth reading. Snippets answer most facts.
• Extract full pages by default. Pass query (--query) only to isolate one fact on a long page; if the passages come back cut off near what you need, extract again without it.
• Rerun with extract_depth set to advanced (--extract-depth advanced) when a page is built with JavaScript, the content is in a table, or basic came back empty. Check failed_results every time and say which pages couldn't be read.

When to stop

• Stop once two independent sources agree on a fact.
• Before you answer, close any gap your draft would mark unconfirmed or fill from memory: run one targeted search or reread the saved page.
• For releases, ownership, official figures and decisions by a named body, confirm on the official domain or repository. A secondary site or a generated answer is not enough.

Hand off the job

When the user wants an in-depth report that needs many sources gathered and weighed, such as an analysis, a comparison across several criteria or an in-depth investigation, use tavily-research instead of your own searches. If it isn't installed, run npx skills add tavily-ai/skills --skill tavily-research and read its SKILL.md first. A single fact or answer to find stays here.