Tavily MCP MCP — 使用場景, 安裝 & 即時演示

為什麼要用

核心特性

四種 MCP 工具：搜尋、擷取、映射、爬行 — LLM 格式的結果
生產級延遲和可靠性（不是抓取駭客）
傳回每個片段的來源 URL — 預設可引用的答案
免費方案（每月 1,000 次 API 呼叫）；付費等級可提高交易量
獨立工作，無需瀏覽器，無需代理，無需反機器人舞蹈

即時演示

實際使用效果

tavily-mcp.replay ▶ 就緒

0/0

安裝

選擇你的客戶端

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "tavily-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "tvly-..."
      }
    }
  }
}

開啟 Claude Desktop → Settings → Developer → Edit Config。儲存後重啟應用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "tavily-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "tvly-..."
      }
    }
  }
}

Cursor 使用與 Claude Desktop 相同的 mcpServers 格式。專案級設定優先於全域。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "tavily-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "tvly-..."
      }
    }
  }
}

點擊 Cline 側欄中的 MCP Servers 圖示，然後選 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "tavily-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "tvly-..."
      }
    }
  }
}

格式與 Claude Desktop 相同。重啟 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "tavily-mcp",
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ]
    }
  ]
}

Continue 使用伺服器物件陣列，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "tavily-mcp": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "tavily-mcp@latest"
        ]
      }
    }
  }
}

加入 context_servers。Zed 儲存後熱重載。

claude mcp add tavily-mcp -- npx -y tavily-mcp@latest

一行命令搞定。用 claude mcp list 驗證，claude mcp remove 移除。

使用場景

實戰用法： Tavily MCP

回答有關模型截止後發生的事情的問題

👤 任何使用 Claude 提出需要新鮮資訊的問題的人 ⏱ ~5 min beginner

何時使用： 您詢問 2026 年版本、最近的 CVE、新的定價頁面或當今的市場 - 模型不知道，必須去查看。

前置條件

泰維利 API 金鑰 — 在 tavilly.com 上註冊（免費套餐 = 1,000 次通話/月）
泰利MCP已安裝 — Paste the config block above into your client's MCP settings

步驟

Ask directly

What did Anthropic ship in Claude Sonnet 4.7 this month? Use Tavily to find the announcement and summarize with sources.✓ 已複製

→ Agent calls tavily_search, returns a summary with linked sources
Drill in on one source

The second source looks most authoritative — use tavily_extract to pull its full text and quote the exact line about context window.✓ 已複製

→ Direct quote with URL + paragraph number

結果： 目前，一輪引用的答案 - 無需手動谷歌搜尋。

注意事項

Search returned SEO junk first — Add site hints: '... from anthropic.com or anthropic's official blog'
Summaries drift from sources — Require direct quotes — 'paraphrase but preserve numbers, dates, names exactly'

搭配使用： filesystem · memory

Do a competitive product scan in one session

👤 PMs, founders, marketers ⏱ ~30 min intermediate

何時使用： You need a one-page brief on every competitor for a given feature category by end of day.

步驟

Discover competitors

Using Tavily, find the top 8 products competing with us in 'AI-native CRM for SMB'. For each, return name, URL, one-liner, founding year.✓ 已複製

→ Structured 8-row table with source links
Map each site

For each competitor, tavily_map their site to find their pricing and features pages. Return the URLs.✓ 已複製

→ 2 URLs per competitor
Extract pricing

tavily_extract each pricing page and build a comparison grid: plan name, monthly price, top 3 differentiators.✓ 已複製

→ Clean grid; cells cite the pricing page URL

結果： A sharable brief with sources — ready for a PMM slide in 30 minutes.

注意事項

Pricing JS-rendered and extract misses it — Fall back to tavily_crawl with render=on, or hit the /pricing sitemap directly

搭配使用： filesystem

Write a tutorial with live-verified links

👤 Technical writers, DevRel ⏱ ~25 min intermediate

何時使用： You're publishing a how-to and every external link has to resolve to the right content today.

步驟

Collect candidate refs

Using tavily_search, find the top 5 canonical docs pages for 'OAuth 2.1 PKCE flow'. Prefer RFCs and vendor docs over blogs.✓ 已複製

→ 5 URLs with a short rationale each
Verify each one

tavily_extract each URL. For each, confirm the page still covers PKCE and flag any that look redirected or stale.✓ 已複製

→ Per-URL live verdict
Embed in the draft

Rewrite my draft tutorial to cite only the verified URLs, with anchor text that matches the page's actual heading.✓ 已複製

→ Updated draft; every link text matches the real page heading

結果： Published tutorial with zero dead links and accurate anchor text.

搭配使用： filesystem

組合

與其他 MCP 搭配，撬動十倍槓桿

tavily-mcp + filesystem

Search, extract into disk, then analyze locally without re-fetching

Search Tavily for recent OWASP top-10 sources, extract them, save to /research/owasp/, then compare the content offline.✓ 已複製

tavily-mcp + memory

Build a research journal that persists between sessions

For each Tavily search, save a one-line note and the URLs to memory under 'project:acme'. Next session, reuse.✓ 已複製

tavily-mcp + context7

Tavily for web context + Context7 for library docs — don't confuse them

Use Context7 for docs questions; Tavily for news, blog posts, and anything not in library indexes.✓ 已複製

工具

此 MCP 暴露的能力

工具	輸入參數	何時呼叫	成本
tavily_search	query: str, max_results?: int, search_depth?: 'basic'\|'advanced', include_domains?: str[]	Primary tool — one query, LLM-ready snippets with URLs	1 API call
tavily_extract	urls: str[], extract_depth?: 'basic'\|'advanced'	You already have a URL and want clean text — no HTML, no ads	1 API call per URL
tavily_map	url: str, max_depth?: int, categories?: str[]	Discover a site's structure — useful before extract/crawl	1 API call
tavily_crawl	url: str, max_depth?: int, limit?: int, instructions?: str	Broad ingest of a small site or doc section — expensive, prefer extract when you already know the URLs	Multiple API calls (one per page)

成本與限制

運行它的成本

API 配額: Free tier = 1,000 API calls/month; scaling plans from $30/mo
每次呼叫 Token 數: Returns ~500–5000 tokens of clean content — much less than raw HTML would
費用: Free tier covers individual daily use; heavy workflows need paid
提示: Prefer tavily_search over tavily_crawl — search is one call, crawl is N. Only crawl when you truly need breadth.

安全

權限、密鑰、影響範圍

憑證儲存： TAVILY_API_KEY in env var (set in the MCP config's env block)

資料出站： Queries and URLs you pass are sent to api.tavily.com. Don't paste proprietary info into the query string.

API key goes to Anthropic's Claude Desktop config in plaintext. Use a key scoped to this use case; rotate if the config file is shared.
tavily_crawl can rack up calls fast — set a low limit unless you know the site size.

故障排查

常見錯誤與修復

401 Unauthorized

Double-check TAVILY_API_KEY in your MCP config. The env block lives inside the server config, not at top level.

驗證： Call any Tavily tool; if the error persists, rotate the key in tavily.com dashboard

Empty results despite a real query

Switch search_depth from 'basic' to 'advanced' for niche topics; add include_domains to bias toward authoritative sources

驗證： Repeat with search_depth: 'advanced'

tavily_extract returns paywalled gibberish

Tavily follows robots.txt and respects paywalls. For paywalled content, note it's unreachable — don't try to bypass.

429 Rate limit

Free tier = 60 RPM. Space out calls, or upgrade at tavily.com. The MCP auto-backs-off once, then surfaces the error.

驗證： Check usage in tavily.com dashboard

替代方案

Tavily MCP 對比其他方案

替代方案	何時用它替代	權衡
exa-mcp-server	You want Exa's neural/embedding-style search and similarity	Different result quality on different query types; both are LLM-tuned
fetch	You have the URL and just want HTML → markdown, no search	No search; no structured extract; you handle the URL discovery
perplexity-ask	You want a research-level answer synthesized by Perplexity, not raw results	Higher latency, higher cost, less composable — it's one big tool
firecrawl-mcp-server	You need heavy crawling or JS rendering as a core workflow	Pricier; overkill for one-off lookups

Tavily MCP

為什麼要用

核心特性

即時演示

實際使用效果

安裝

選擇你的客戶端

使用場景

實戰用法： Tavily MCP

回答有關模型截止後發生的事情的問題

前置條件

步驟

注意事項

Do a competitive product scan in one session

步驟

注意事項

Write a tutorial with live-verified links

步驟

組合

與其他 MCP 搭配，撬動十倍槓桿

工具

此 MCP 暴露的能力

成本與限制

運行它的成本

安全

權限、密鑰、影響範圍

故障排查

常見錯誤與修復

替代方案

Tavily MCP 對比其他方案

更多

資源