MCP servers / Tavily

Tavily MCP server: setup for Claude Code, Claude Desktop and Cursor

Tavily is a search service built for AI assistants, and this server connects your AI to it. It has five tools: search, extract the content of URLs, crawl a website, map a site's structure, and a research tool for broader questions.

It overlaps with Brave Search for finding pages and with Firecrawl for reading them, in a smaller package.

Runs with

npx (Node.js)

Package

tavily-mcp on npm

Needs

Node.js (for npx) and a Tavily API key from tavily.com.

Tools

5 (not declared as read-only or not)

Source

tavily-ai/tavily-mcp

Last checked

2026-09-20 (server reported version 0.2.22)

Config for Tavily

Generated by the same code as the free builder, so it matches what you get there. Pick your client, then copy.

Claude Code

Save as .mcp.json in your project root. Claude Code asks you to approve it the first time.

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

Set TAVILY_API_KEY in your environment before starting Claude Code. The config only references the variable, so it is safe to commit.

Or add it from a terminal instead (single quotes work in bash, zsh and PowerShell; more on this command):

claude mcp add tavily -e 'TAVILY_API_KEY=${TAVILY_API_KEY}' -- npx -y tavily-mcp@latest

Claude Desktop

Merge into claude_desktop_config.json (Settings → Developer → Edit Config), then fully quit and restart Claude Desktop.

{
  "mcpServers": {
    "tavily": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "paste-your-key-here"
      }
    }
  }
}

Claude Desktop keeps TAVILY_API_KEY in this file as plain text. Keep the file private and never share or commit it.

Cursor

Save as .cursor/mcp.json in your project, or ~/.cursor/mcp.json for all projects.

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

Set TAVILY_API_KEY in your environment before launching Cursor. The config only references the variable, so it is safe to commit.

What it can do

These are all 5 tools the server reported when we started it and asked on 2026-09-20. This server does not say which tools are read-only, so we sorted them by what each name and description says.

• tavily_search reads (by name)
  Search the web for current information on any topic.
• tavily_extract reads (by name)
  Extract content from URLs.
• tavily_crawl reads (by name)
  Crawl a website starting from a URL.
• tavily_map reads (by name)
  Map a website's structure.
• tavily_research reads (by name)
  Perform comprehensive research on a given topic or question.

Watch out for

• Requests run on Tavily's service under your account, so check your plan's limits before letting the AI crawl a site or run broad research.
• None of the five tools declares itself read-only or not. They search and fetch rather than change anything on your machine, so we treated them as reads.
• The README also documents a hosted remote server, with a sign-in flow, as an alternative to running this one locally. We only tested the local npx package.
• The README lets you set defaults for searches through a DEFAULT_PARAMETERS environment variable holding a JSON object (see below). We confirmed the server starts with it set. We did not test how each option changes results.

Variations

Set default search options

A JSON object of defaults for the search tool, as in the README's example. Here it caps results at five. The value is not a secret, so it is written directly in the config.

{
  "mcpServers": {
    "tavily": {
      "command": "npx",
      "args": [
        "-y",
        "tavily-mcp@latest"
      ],
      "env": {
        "TAVILY_API_KEY": "${TAVILY_API_KEY}",
        "DEFAULT_PARAMETERS": "{\"max_results\": 5}"
      }
    }
  }
}

Shown in Claude Code format. Cursor uses ${env:NAME} for keys, and Claude Desktop takes the key itself, exactly as in the configs above.

Try asking

• “Search for the current best practices for rate limiting an API and give me the top three sources.”
• “Extract the main content from this documentation page and summarize the setup steps.”
• “Map the structure of this docs site so I can see its main sections.”

Related servers

• Brave Search: Search the web, news, images, videos and local places with the Brave Search API.
• Firecrawl: Scrape and crawl websites into clean text.