Tavily MCP ServerFeatured

What is Tavily MCP Server?

Tavily MCP Server is the official Tavily MCP integration that gives AI assistants real-time web search, page extraction, site mapping, crawling and multi-source research through the Tavily API. You can use it as a hosted remote server or run it locally with npx.

The project is maintained by Tavily (tavily-ai/tavily-mcp, MIT license) and works with Claude Code, Claude Desktop, Cursor and any other MCP client. The hosted endpoint supports an API key in the URL, a Bearer header, or OAuth sign-in.

How to use Tavily MCP Server?

You need a Tavily API key from app.tavily.com. Tavily offers a free account, and its docs describe the free key as 1,000 credits per month with no credit card. Keys use the tvly- prefix. For the local server you also need Node.js 20 or higher.

Remote (hosted) URL — nothing to install:

https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>

If your client can send headers, you can drop the key from the URL and send Authorization: Bearer <your-api-key> instead. You can also connect to https://mcp.tavily.com/mcp with no key and sign in through OAuth.

Local (stdio):

npx -y tavily-mcp@latest

with TAVILY_API_KEY set in the environment.

Claude Code

Add the hosted Tavily MCP server with an API key in the URL:

claude mcp add --transport http tavily https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>

Or add it without a key and use OAuth:

claude mcp add --transport http tavily https://mcp.tavily.com/mcp

Then run claude, type /mcp, select the Tavily server and finish the sign-in. Add --scope user to either command to make it available in all your projects.

Claude Desktop

Add the local server to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

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

Tavily's docs also describe adding the remote URL as an integration in Claude Desktop's settings.

Cursor

Use the "Add to Cursor" button in the README, or add this to mcp.json and replace the key:

{
  "mcpServers": {
    "tavily-remote-mcp": {
      "command": "npx -y mcp-remote https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>",
      "env": {}
    }
  }
}

For OAuth instead, use npx mcp-remote https://mcp.tavily.com/mcp as the command.

Other clients

Clients that only support local stdio servers can reach the hosted server through the mcp-remote bridge:

{
  "tavily-remote": {
    "command": "npx",
    "args": ["-y", "mcp-remote", "https://mcp.tavily.com/mcp/?tavilyApiKey=<your-api-key>"]
  }
}

Default parameters: set the DEFAULT_PARAMETERS env var locally, or send a DEFAULT_PARAMETERS header to the remote server, containing JSON such as {"include_images": true, "max_results": 15, "search_depth": "advanced"}.

Key features of Tavily MCP Server

• Official server built and maintained by Tavily, MIT licensed, published on npm as tavily-mcp.
• Hosted remote endpoint at mcp.tavily.com, so you can skip a local install.
• Three ways to authenticate on the remote server: API key in the query string, a Bearer header, or OAuth.
• Search options for depth (basic, advanced, fast, ultra-fast), topic (general or news), time range, start and end dates, and domains to include or exclude.
• Extraction returns page content as markdown or text, ready for an LLM to read.
• Site mapping and crawling with configurable depth and breadth.
• A research tool that pulls together findings from many sources into a single detailed answer.
• DEFAULT_PARAMETERS sets search defaults once, and the optional TAVILY_HUMAN_ID forwards an end-user ID for analytics. Tavily hashes that ID before storing it.

Tavily MCP Server tools

The tools fall into three groups:

• Search: tavily_search runs a real-time web search and returns snippets with source URLs. It covers news, recent events and facts beyond the model's training cutoff.
• Reading websites: tavily_extract returns the content of one or more URLs. tavily_map lists the URLs found on a site, starting from a base URL. tavily_crawl explores a site from a starting URL and extracts the content of the pages it visits.
• Research: tavily_research gathers information from multiple sources for a broad question and returns a detailed response. Its tool description sets a limit of 20 requests per minute.

Use cases of Tavily MCP Server

• Current events and fact checks: "Search for news articles about AI startups from the last 7 days."
• Searching specific sites: "Search for climate change research on nature.com and sciencedirect.com."
• Reading an article: "Extract the main content from https://example.com/article and summarize it."
• Docs lookups in a coding assistant: connect Tavily in Claude Code or Cursor so the agent can check current library docs while it writes code.
• Auditing a site's structure: "Map docs.example.com and list every page under /guides."
• Research reports: "Research the current state of solid-state batteries and give me a sourced summary."

Tavily MCP Server vs alternatives

• Exa MCP Server is another search API built for AI agents. Choose based on result quality for your queries and on pricing.
• Firecrawl MCP Server focuses on scraping and crawling. Tavily leads with search and also includes extraction, mapping and crawling.
• Puppeteer MCP Server and Playwright MCP Server drive a real browser. Use them when a page needs clicks, logins or form input instead of a plain fetch.

Several community "Tavily MCP" forks exist. This listing points to the official tavily-ai repository.

Troubleshooting Tavily MCP Server

• Server not found or fails to start: run node --version and npm --version to confirm Node.js 20 or higher, and check your config file for JSON syntax errors.
• npx not found from the desktop app: run which npx and put the full path to npx in the command field.
• Invalid API key or usage limit errors: make sure the key starts with tvly- and is active in your Tavily dashboard. You can test it in the Tavily Playground.
• OAuth stuck or wrong account: run rm -rf ~/.mcp-auth to clear stored credentials, then sign in again. To control which key OAuth uses, name it mcp_auth_default in your Tavily account.
• Tool names look different: older docs write tavily-search and tavily-extract. Current releases expose tavily_search, tavily_extract and the other tools with underscores.