Tavily MCP Server

Executes semantic web searches via the Tavily API and returns structured results with relevance scoring, source attribution, and clean text extraction. The MCP server acts as a bridge that translates search queries into Tavily API calls, handling authentication via environment variables or URL parameters, and formats responses as JSON with ranked results including URLs, snippets, and confidence scores. Results are pre-processed to remove boilerplate and optimize token efficiency for LLM consumption.

Extracts and cleans full-page content from specified URLs, returning structured text with semantic understanding of page layout and content hierarchy. The tavily-extract tool uses Tavily's content extraction engine to parse HTML, remove navigation/ads/boilerplate, and return clean markdown or plain text. It handles authentication via the same MCP transport layer and returns metadata including extraction confidence and source attribution.

Executes autonomous research workflows that combine search, extraction, and analysis in a single MCP tool call. The tavily-research tool accepts a research query and automatically performs multiple search iterations, extracts content from promising sources, and synthesizes findings into a structured research report. This tool orchestrates the search and extract capabilities internally, handling retry logic and source validation without requiring the client to manually chain multiple tool calls.

Crawls websites starting from a seed URL and discovers linked pages, returning a structured map of the site's content hierarchy. The tavily-crawl tool uses Tavily's crawler to traverse links, respect robots.txt, and extract metadata from discovered pages. Results include page URLs, titles, content snippets, and relationship information (parent/child links), enabling clients to understand site structure without manual link parsing.

Generates a semantic map of a website's content by crawling and categorizing pages based on topic, content type, and relevance. The tavily-map tool combines crawling with NLP-based content analysis to produce a hierarchical map showing how pages relate to each other conceptually, not just structurally. Results include topic clusters, content type distribution, and recommended navigation paths.

Implements the Model Context Protocol (MCP) server specification using TypeScript and Node.js, handling bidirectional communication with MCP clients via standard input/output (stdio). The server instantiates an MCP Server instance, registers the five Tavily tools as callable handlers, and uses StdioServerTransport to manage message serialization/deserialization. Tool handlers are registered via setRequestHandler(ListToolsRequestSchema, ...) and CallToolRequestSchema, mapping incoming MCP requests to Tavily API calls and returning structured responses.

Supports both remote (cloud-hosted at https://mcp.tavily.com/mcp/) and local (self-hosted via NPX, Docker, or Git) deployment models, with identical tool capabilities but different authentication and infrastructure patterns. Remote deployment uses URL parameters or Bearer token headers for authentication and requires no local setup. Local deployment uses environment variables for API keys and can be containerized with Docker or run directly via NPX. Both models expose the same five tools through the MCP protocol.

Provides native integration with multiple MCP-compatible clients through configuration files and environment setup. For Claude Desktop, the server is configured via claude_desktop_config.json with command and arguments. For Cursor and VS Code, integration uses MCP settings in client configuration. For OpenAI, the server bridges via mcp-remote (a separate tool that exposes MCP servers as OpenAI function-calling APIs). Each integration method handles authentication, tool discovery, and response formatting differently based on the client's capabilities.

Handles Tavily API key authentication through multiple mechanisms depending on deployment model. For local deployment, the server reads TAVILY_API_KEY from environment variables at startup and uses it for all Tavily API calls via axios instance configuration. For remote deployment, clients can pass the API key as a URL query parameter (?tavilyApiKey=...) or as an Authorization header (Bearer token). The server validates the API key on each request and returns 401 errors for invalid/missing keys.

Formats all tool responses as structured JSON with consistent fields for relevance scoring, source attribution, and content metadata. Each search result includes a relevance_score (0-1), source URL, and confidence metrics. Extract responses include extraction_confidence_score and source_attribution. Research responses include sources_used array with confidence scores for each source. This standardized formatting enables clients to programmatically rank, filter, and cite sources without parsing unstructured text.