Tavily MCP: Web Search, Extract and Crawl for Claude
Tavily's official MCP server gives Claude real-time web search, page extraction, site mapping and crawling. Claude Code commands, OAuth or API key, caveats.
By Claudekit Team Published 2 min read
- Type
- MCP Server
- Pricing
- Freemium
- Repository
- github.com
- GitHub stars
- 2,454
- Website
- docs.tavily.com
The Tavily MCP server is Tavily’s official Model Context Protocol server for web search. It gives Claude four tools: real-time search, content extraction from pages, a site mapper and a crawler. Use it when Claude needs current information from the web, with sources it can cite.
Who it is for
Developers and researchers who want Claude Code to look up current docs, changelogs or news, or to pull clean text from a set of pages, without writing a scraper.
Install
Tavily hosts a remote server, and its README documents the Claude Code commands. With OAuth (no key in the URL):
claude mcp add --transport http tavily https://mcp.tavily.com/mcpThen run claude, type /mcp, select Tavily and complete the sign-in. Or pass your API key in the URL:
claude mcp add --transport http tavily https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>Add --scope user to make it available in every project. A key in the URL is stored in your MCP config in plain text, so keep that file out of version control. To run locally instead, the package is tavily-mcp on npm with a TAVILY_API_KEY environment variable.
Use cases
- Current facts: “Search for the latest release notes of this library and summarize breaking changes.”
- Page extraction: pull the text of a few docs pages so Claude can compare them.
- Site mapping: map a docs site before deciding which pages to read.
- Research: crawl a section of a site and build a summary with links.
Pros
- Official, MIT licensed and hosted, so no local install is needed.
- Documented Claude Code commands for both OAuth and API key auth.
- Default parameters let you set result count or search depth once.
- Search, extract, map and crawl in one server.
Cons and caveats
- Usage counts against your Tavily plan. Crawls can use many credits.
- Web content is untrusted input. Watch for prompt injection in fetched pages.
- Extracted pages can be long. Ask for summaries to protect your context window.
Links
- Repo: tavily-ai/tavily-mcp
- Docs: Tavily MCP
- Package: tavily-mcp on npm
- Related: Exa MCP server, Brave Search MCP server, Firecrawl MCP server
Checked 24 Oct 2026 against the vendor repo, the Tavily docs and the MCP registry.
Frequently asked questions
Do I need a Tavily API key?
You need a Tavily account. You can pass an API key in the URL or an Authorization header, or connect without a key in the URL and finish an OAuth flow from /mcp in Claude Code.
Which tools does it provide?
Search, extract, map and crawl: tavily-search for real-time results, tavily-extract for page content, plus site mapping and crawling tools.
Can I set default search options?
Yes. Send a DEFAULT_PARAMETERS header (remote) or environment variable (local) with a JSON object such as max_results or search_depth.
What license is it under?
The repository is MIT licensed.