Tavily API

Executes live web searches and returns results ranked and formatted specifically for LLM consumption rather than human browsing. Uses intelligent result filtering to surface relevant content while removing boilerplate, ads, and low-signal pages. Implements search depth controls allowing callers to trade latency for comprehensiveness (shallow vs deep crawl modes). Returns structured, chunked content pre-formatted for token efficiency in LLM context windows.

Restricts web search results to specific domains or domain categories, allowing callers to filter searches to trusted sources, exclude low-quality sites, or focus on particular content types (e.g., academic papers, news sites, documentation). Implements domain filtering at query time rather than post-processing results, reducing wasted API credits on irrelevant sources. Exact filtering syntax and supported domain categories are not documented in public materials.

Integrates with the Model Context Protocol (MCP) standard through a partnership with Databricks, allowing Tavily search to be exposed as an MCP resource that compatible clients (Claude, other MCP-aware applications) can discover and use. Enables standardized, composable tool integration without provider-specific code. Exact MCP schema and resource definitions are not documented.

Provides free tier access with 1,000 API credits per month (no credit card required), allowing developers to prototype and test Tavily integration without upfront costs. Credits reset monthly on an unspecified date. Exact credit-to-operation mapping is not documented, making it unclear how many searches/extractions the free tier supports.

Offers flexible pay-as-you-go pricing at $0.008 per API credit, allowing developers to scale usage without committing to monthly plans. Billing is based on actual usage rather than fixed monthly allocations. Exact credit-to-operation mapping and overage handling are not documented, making cost prediction difficult.

Offers monthly subscription plans bundling 4,000+ API credits per month at fixed prices, providing better per-credit rates than pay-as-you-go pricing for committed usage. Plans include 'Project' tier with adjustable pricing slider and higher rate limits than free tier. Exact pricing, rate limits, and credit-to-operation mapping are not documented.

Offers enterprise tier with custom pricing, custom rate limits, and 99.99% uptime SLA for mission-critical applications. Includes dedicated support and customized integration assistance. Exact SLA terms, support response times, and customization options are not documented.

Extracts and structures content from individual web pages, converting HTML/DOM into clean, LLM-ready text or structured data. Handles boilerplate removal (navigation, ads, footers), text cleaning, and optional content chunking for large pages. Designed as a complement to search — after search identifies relevant URLs, extraction provides deep content access without requiring the caller to parse HTML or manage DOM complexity.

Crawls multiple pages starting from a seed URL, following internal links up to a specified depth limit. Returns aggregated content from crawled pages, useful for extracting information from documentation sites, product catalogs, or research repositories. Implements depth controls to prevent runaway crawls and manage API credit consumption. Exact crawl behavior (link selection strategy, depth definition, concurrent crawl limits) is not documented.

Newly released 'research' endpoint that aggregates and synthesizes information from multiple sources to produce research-grade summaries. Claimed to achieve 'state-of-the-art' results, though specific methodology and model details are not documented. Appears to combine search, extraction, and synthesis to produce comprehensive research outputs rather than simple ranked result lists. Exact input/output format and synthesis approach are unknown from public materials.

Implements security layers to detect and block prompt injection attempts in search queries and to filter out malicious or adversarial content from search results. Prevents attackers from manipulating search queries to extract sensitive information or bypass safety guardrails. Filters results to exclude known malicious sources, phishing sites, and content designed to exploit LLM vulnerabilities. Exact detection mechanisms and filtering criteria are not documented.

Detects and redacts personally identifiable information (PII) in search results before returning them to callers, preventing accidental exposure of sensitive data like email addresses, phone numbers, social security numbers, or financial information. Operates as a post-processing filter on search results. Exact PII detection patterns and redaction approach are not documented.

Caches search results for repeated or similar queries, returning cached results with minimal latency (likely <50ms) instead of executing fresh web searches. Reduces API credit consumption for repeated queries and improves response time for common searches. Cache invalidation strategy and similarity matching algorithm are not documented.

Automatically chunks extracted web content into LLM-friendly segments sized to fit within typical context windows (likely 1-4KB chunks), with metadata about chunk boundaries and source attribution. Removes unnecessary whitespace, normalizes formatting, and optimizes token efficiency to maximize information density per token. Designed to reduce the number of tokens consumed when injecting search results into LLM prompts.

Provides native integrations with major LLM provider APIs (OpenAI, Anthropic, Groq), allowing Tavily search results to be automatically injected into LLM function calls or tool use workflows without custom integration code. Likely implements function calling schemas compatible with each provider's tool-use format, enabling LLMs to call Tavily search as a native tool. Exact integration mechanism and supported function calling patterns are not documented.