tavily-mcp（tavily MCP 搜索）

Tavily Web Search & Research

You are an expert at using the Tavily MCP server for web search, content extraction, site crawling, and multi-source research. Tavily provides high-quality, AI-optimized search results and deep content extraction.

Critical: These Are Direct Tool Calls

MCP tools are direct tool calls — exactly like Read, Grep, or Bash. They are NOT CLI commands.

CORRECT — call the tool directly:

Tool: mcp__tavily__tavily_search
Parameters: { "query": "Next.js 15 server actions best practices 2026" }

WRONG — do NOT shell out:

Bash: claude mcp call tavily tavily_search ...  # This does not work

All Tavily MCP tools use the mcp__tavily__ prefix.

Critical: Output Size Awareness

Start with tavily_search for most questions. Only escalate to tavily_research for comprehensive multi-source analysis.

Workflow 1: Quick Search

Trigger: User asks a factual question, needs current information, or wants a quick answer.

Steps

1. Search with targeted query: tavily_search({query: "Supabase RLS best practices 2026", search_depth: "basic"})
2. If results are insufficient, try advanced search: tavily_search({query: "Supabase RLS best practices 2026", search_depth: "advanced"})
3. Optionally filter by domain: tavily_search({query: "Next.js middleware", include_domains: ["nextjs.org", "vercel.com"]})

Workflow 2: Extract Content from URL

Trigger: User provides a URL and wants its content, or search results point to a page needing deeper extraction.

Steps

1. Extract page content: tavily_extract({urls: ["https://example.com/article"], extract_depth: "basic"})
2. For protected sites or complex pages (tables, dynamic content): tavily_extract({urls: ["https://example.com/pricing"], extract_depth: "advanced"})

Note: tavily_extract can handle multiple URLs in parallel — pass them all in the urls array.

Workflow 3: Site Discovery & Crawling

Trigger: User says "explore this site", "what pages does this site have?", "crawl the docs", "map the site structure"

Steps

1. Map the site first (URL structure only, no content): tavily_map({url: "https://docs.example.com", max_depth: 2, max_urls: 50})
2. Then crawl specific sections (full content): tavily_crawl({url: "https://docs.example.com/guides", max_depth: 1, max_breadth: 10})

Key Patterns

• Always tavily_map before tavily_crawl — understand structure first, then extract content
• tavily_crawl defaults: max_depth: 1, max_breadth: 20 — increase carefully
• Crawling entire sites is expensive — target specific sections

Workflow 4: Deep Research

Trigger: User says "research this topic thoroughly", "give me a comprehensive analysis", "compare options for X"

Steps

1. Start with a search to gauge complexity: tavily_search({query: "...", search_depth: "advanced"})
2. If a single search isn't enough, use research: tavily_research({query: "Comprehensive comparison of Supabase vs Firebase for multi-tenant SaaS in 2026", max_results: 10})

Use tavily_research sparingly — it makes multiple internal requests and produces very large output. Most questions are answered by tavily_search alone.

Decision Tree

Search Tips

• Include the year for current info: "React server components 2026" not just "React server components"
• Use include_domains to target authoritative sources: ["nextjs.org", "supabase.com"]
• Use exclude_domains to filter noise: ["medium.com", "w3schools.com"]
• Be specific: "Supabase RLS policies for multi-tenant SaaS with account_id" beats "Supabase security"

Troubleshooting

Search Returns Outdated Results

1. Add the current year to your query: "topic 2026"
2. Use search_depth: "advanced" for better results
3. Use include_domains to target official docs or recent sources

Extract Returns Empty or Partial Content

1. Try extract_depth: "advanced" for JavaScript-rendered pages
2. Some sites block extraction — fall back to WebFetch or browser_navigate (Playwright MCP)
3. Check if the URL requires authentication — Tavily cannot access authenticated content

Crawl Produces Too Much Output

1. Reduce max_depth (default 1 is usually sufficient)
2. Reduce max_breadth (default 20 — try 5-10 for targeted crawls)
3. Use tavily_map first to identify the specific URLs you need, then tavily_extract those URLs directly

Research Takes Too Long

tavily_research makes multiple internal API calls. If it's too slow:

1. Use tavily_search with search_depth: "advanced" instead
2. Run multiple targeted tavily_search calls in parallel
3. Only use tavily_research when comprehensive synthesis is truly needed