CtrlK
See also: Retrieval-augmented generation, AI agent, and Information retrieval
Tavily is a search engine and API platform built specifically for AI agents and retrieval-augmented generation (RAG) systems. Unlike traditional search engines designed for human users, Tavily optimizes every part of its pipeline for consumption by large language models (LLMs), returning structured, ranked, and relevance-scored results that can be fed directly into AI workflows. The platform provides a suite of APIs, including Search, Extract, Map, and Crawl endpoints, that allow developers to connect their AI applications to real-time web data through a single integration [1][2].
Tavily was founded in late 2023 by Rotem Weiss and Assaf Elovic and is headquartered in New York City. The company grew out of an open-source project called GPT Researcher, which Elovic created in 2023 to fetch real-time web data for LLMs before ChatGPT had internet connectivity [3]. By August 2025, Tavily had raised $25 million in total funding, including a $20 million Series A led by Insight Partners and Alpha Wave Global [3]. In February 2026, Nebius, an AI infrastructure company, announced an agreement to acquire Tavily for an initial $275 million, with the deal potentially rising to $400 million based on performance milestones [4][5].
Tavily serves Fortune 500 enterprises and leading AI companies, including IBM, Amazon Web Services, MongoDB, Cohere, Groq, JetBrains, Monday.com, and LangChain [5][6]. It integrates natively with popular AI frameworks such as LangChain, LlamaIndex, and supports the Model Context Protocol (MCP) for direct connectivity with AI assistants [7][8].
Tavily traces its roots to an open-source project called GPT Researcher, created by Assaf Elovic in 2023. At the time, ChatGPT and other consumer-facing LLMs lacked the ability to access live web data. Elovic, a software engineer who had previously served as Head of R&D at Wix and founded Tiv.ai, built GPT Researcher as an autonomous agent that could conduct deep research on any topic by searching the web, aggregating information from multiple sources, and generating comprehensive reports [9][3].
The project uses a planner-executor architecture: planner agents generate research questions, execution agents gather relevant information from across the web, and a publisher agent aggregates the findings into a structured report. An average research task takes roughly three minutes and costs about $0.10 in API calls. GPT Researcher supports multiple output formats, including PDF, Word, and Markdown, and can work with various LLM providers such as OpenAI, Claude, Gemini, Ollama, and Mistral [9].
GPT Researcher gained significant traction in the developer community. As of early 2026, the project has accumulated over 26,000 GitHub stars and more than 3,500 forks. It is licensed under the Apache 2.0 license [9].
The viral success of GPT Researcher led Elovic to team up with Rotem Weiss, a data scientist and Columbia University graduate, to commercialize the underlying technology. In late 2023, they founded Tavily with the goal of building a search engine purpose-built for AI agents and LLM applications [3][10].
Yuval Rozio, Managing Director at Alpha Wave Global, also played a role in the company's formation, as Alpha Wave Global incubated the startup during its early stages [10].
Tavily raised a $5 million seed round before securing its Series A. In August 2025, Insight Partners led a $20 million Series A funding round, with participation from Alpha Wave Global. This brought the company's total funding to $25 million. At the time of the Series A, Tavily was roughly one year old and already generating significant revenue from a mix of AI startups and Fortune 500 enterprises [3].
On February 10, 2026, Nebius, an AI infrastructure company that spun out of Russian search giant Yandex in February 2023, announced an agreement to acquire Tavily. Bloomberg reported the initial deal value at $275 million in cash, with the total potentially rising to $400 million if certain milestones are met [4][5]. Nebius operates premium data centers providing cloud compute for AI workloads and sought to add agentic search capabilities to its platform. Roman Chernin, Nebius co-founder, stated that the acquisition "brings the search layer directly into our stack, so developers can focus on their applications instead of managing multiple vendors" [5].
The acquisition occurred in the context of rapid growth in the agentic AI market, which is projected to expand from approximately $7 billion in 2025 to between $140 billion and $200 billion by the early 2030s [5].
Tavily offers a suite of APIs designed to give AI agents and RAG systems structured access to real-time web data. All endpoints are authenticated using API keys and return results in JSON format optimized for LLM consumption [1][2].
The Search API is Tavily's core product. It performs semantic search optimized for AI workflows, returning concise, ranked snippets designed for downstream reasoning and retrieval-augmented generation. Unlike traditional search APIs that return raw URLs and brief snippets, Tavily aggregates up to 20 sites per request, uses proprietary AI to score and filter sources, and returns structured results with titles, content snippets, source URLs, and relevance scores [1][11].
The Search API supports four depth levels that trade off between latency and relevance:
| Search Depth | Description | Credit Cost |
|---|---|---|
| Ultra-fast | Minimizes latency for time-critical use cases | 1 credit |
| Fast | Lower latency with good relevance | 1 credit |
| Basic | Balanced relevance and latency for general-purpose searches; returns one NLP summary per URL | 1 credit |
| Advanced | Highest relevance with increased latency; returns multiple semantically relevant snippets per URL | 2 credits |
The API also supports topic categories (general, news, and finance), temporal filtering (by day, week, month, year, or specific date ranges), domain inclusion and exclusion lists, country-based result boosting, safe search filtering, and an auto-parameters mode that automatically configures search settings based on query intent [11].
When the include_answer parameter is enabled, Tavily generates a short LLM-powered answer alongside the search results, which is useful for agent-to-agent communication. The include_raw_content option returns the full parsed content of each page in Markdown or plain text format [11].
The Extract API retrieves raw content from a list of specified URLs, making it suitable for data collection, content analysis, and research tasks. It converts web pages into clean, structured text that LLMs can process directly. The API supports two extraction depth levels: basic and advanced. Users can choose between Markdown and plain text output formats, and can optionally include images and favicons in the response [12].
Extract costs 1 credit per 5 successful URL extractions at basic depth, or 2 credits per 5 URLs at advanced depth [13].
The Map API discovers and visualizes the structure of a website starting from a base URL. It returns a list of pages and their relationships, allowing developers to understand a site's layout before extracting content. Map costs 1 credit per 10 pages discovered, or 2 credits per 10 pages when custom instructions are provided [12][13].
The Crawl API automatically explores and extracts content from an entire website by providing a single starting URL. It traverses a site like a graph, following internal links and gathering content from each page it visits. Developers can control the crawl's depth, breadth, and total page limit. Additional parameters include natural language queries to prioritize specific content, regex-based path and domain filters, the option to follow external links, and category filters for targeting specific page types such as documentation or blog posts [14].
Crawl combines the costs of mapping and extraction, so the total credit usage depends on the number of pages discovered and the extraction depth selected [13].
The Research API provides an automated deep research capability built on top of the other APIs. It generates comprehensive research reports on any topic by coordinating multiple search, extraction, and synthesis steps. Two model tiers are available: the "pro" model costs between 15 and 250 credits per request, while the "mini" model costs between 4 and 110 credits per request [13].
Tavily's search pipeline is designed to transform raw web data into LLM-ready content in a single API call. When a query is submitted, the system performs the following steps [1][2]:
The architecture incorporates dynamic caching and an agent-native index to maintain low latency even at production scale [2].
Tavily positions itself as a firewall between AI agents and the open internet. All retrieved content passes through security, privacy, and content validation layers that are designed to block several categories of threats [2][15]:
Tavily holds SOC 2 certification and operates with a zero data retention policy, meaning that search queries and results are not stored after processing. The company also partnered with Pillar Security to provide additional enterprise-grade security guardrails for AI web access [15].
The Tavily Python SDK (tavily-python) provides a full-featured client library for interacting with all Tavily API endpoints. It supports synchronous and asynchronous operations (async support was added in version 0.3.4), and covers the complete range of search, extract, crawl, map, and research functionalities. The package is available on PyPI [16].
The official JavaScript SDK (@tavily/core) offers the same capabilities as the Python SDK for Node.js and browser environments. A separate package (@tavily/ai-sdk) provides compatibility with Vercel's AI SDK (versions 5 and 6) for building AI-powered web applications [16].
All Tavily functionality is accessible through a RESTful API at https://api.tavily.com/. Authentication is handled via Bearer tokens in the HTTP Authorization header. The API returns JSON responses and supports standard HTTP error codes for invalid requests (400), authentication failures (401), rate limiting (429), plan limit exceedance (432), and server errors (500) [11].
Tavily provides a Model Context Protocol (MCP) server that connects AI assistants such as Claude, Cursor, and other MCP-compatible tools directly to Tavily's search, extract, map, and crawl capabilities. The MCP server is available both as a remote hosted service (at https://mcp.tavily.com/mcp/) and as a locally installable package. Authentication can be handled through OAuth or API key query parameters [8].
Tavily has a dedicated LangChain integration package (langchain-tavily for Python, @langchain/tavily for JavaScript). This integration provides tool classes that can be bound directly to LangChain agents, giving them the ability to perform real-time web searches. The tool accepts customizable parameters including max_results, search_depth, topic, time_range, include_domains, exclude_domains, and more. Tavily is one of the default search tools recommended in LangChain's agent documentation [7].
The TavilyToolSpec for LlamaIndex allows developers to integrate Tavily's search and research capabilities into LlamaIndex agent workflows. The integration is available through the llama-index-tools-tavily-research package on PyPI and is listed on LlamaHub [17].
While CrewAI does not include a native Tavily integration out of the box, any LangChain or LlamaIndex tool can be converted into a CrewAI tool. This allows CrewAI multi-agent systems to use Tavily for web search and content extraction tasks [18].
The @tavily/ai-sdk package provides direct compatibility with Vercel's AI SDK, enabling developers building Next.js or other React-based AI applications to integrate real-time web search with minimal setup [16].
Tavily offers a tiered pricing structure with monthly subscription plans and a pay-as-you-go option. All plans include access to the Search, Extract, Map, and Crawl APIs. Credits do not roll over between billing periods [13].
| Plan | Monthly Credits | Price | Per-Credit Cost |
|---|---|---|---|
| Researcher (Free) | 1,000 | $0 | Free |
| Project | 4,000 | $30/month | $0.0075 |
| Bootstrap | 15,000 | $100/month | $0.0067 |
| Startup | 38,000 | $220/month | $0.0058 |
| Growth | 100,000 | $500/month | $0.005 |
| Pay-as-you-go | Variable | Per usage | $0.008 per credit |
| Enterprise | Custom | Custom | Custom |
The free Researcher plan provides 1,000 API credits per month with no credit card required, making it accessible for individual developers and small experiments. Enterprise plans offer custom credit volumes, dedicated support, and enhanced security features [13].
| Operation | Basic Depth | Advanced Depth |
|---|---|---|
| Search | 1 credit per request | 2 credits per request |
| Extract | 1 credit per 5 URLs | 2 credits per 5 URLs |
| Map | 1 credit per 10 pages | 2 credits per 10 pages (with instructions) |
| Research (Pro) | 15-250 credits per request | N/A |
| Research (Mini) | 4-110 credits per request | N/A |
Tavily's primary use case is providing the retrieval layer for RAG systems. Instead of building and maintaining custom web scraping infrastructure, developers can use the Search API to fetch relevant, up-to-date information that supplements the knowledge of their LLMs. The structured JSON output is designed to be inserted directly into LLM prompts as context, reducing the need for post-processing [1][2].
For autonomous AI agents, Tavily serves as the interface between the agent and the live web. Agents can call the Search API to answer factual questions, the Extract API to read specific web pages, and the Crawl API to index entire documentation sites. This is particularly valuable for agents performing tasks like lead research in sales, fraud detection in finance, and customer support automation [3][6].
The Research API and the open-source GPT Researcher project enable fully automated research workflows. Given a topic, the system generates research questions, searches the web for relevant information, synthesizes findings, and produces a multi-page report with citations. This approach is used for market research, competitive analysis, academic literature reviews, and content creation [9].
The Crawl API enables organizations to build and maintain knowledge bases by systematically extracting content from documentation sites, wikis, and other structured web sources. Extracted content can be stored in vector databases for semantic search or used to populate internal knowledge management systems [14].
AI chatbots use Tavily to ground their responses in current, verifiable information. By calling the Search API before generating a response, chatbots can provide answers that include source citations and reflect the latest available information rather than relying solely on their training data [1].
Tavily operates in the growing market for AI-native search APIs. Several competitors offer overlapping capabilities with different strengths and trade-offs.
| Feature | Tavily | Exa | SerpAPI |
|---|---|---|---|
| Search approach | AI-ranked aggregation from multiple sources | Semantic search over proprietary index | Raw access to Google and 20+ search engines |
| Optimized for | LLMs and AI agents | Deep research and content retrieval | SEO monitoring and SERP data |
| Free tier | 1,000 credits/month | 2,000 one-time credits | 250 searches/month |
| Content extraction | Built-in Extract and Crawl APIs | HTML cleaning and content parsing | Structured SERP data only |
| Response speed (p95) | 3.8s - 4.5s | 1.4s - 1.7s | Varies by engine |
| People/company search | Not available | Indexes 1B+ LinkedIn profiles | Via Google/Bing results |
| LLM framework integrations | LangChain, LlamaIndex, MCP, Vercel AI SDK | LangChain, LlamaIndex | LangChain |
Other competitors include WebSearchAPI, Firecrawl (focused on web scraping and crawling), Perplexity's Sonar API (focused on conversational search), and Serper (a lightweight Google SERP API) [19].
GPT Researcher remains Tavily's flagship open-source project and serves as both a showcase for Tavily's API capabilities and a standalone research tool. The project is maintained at github.com/assafelovic/gpt-researcher under the Apache 2.0 license [9].
GPT Researcher uses a multi-agent architecture inspired by the STORM paper from Stanford University. The system coordinates three types of agents [9]:
As of early 2026, GPT Researcher offers the following features [9]:
The backend is built with Python and FastAPI, with LangGraph support for multi-agent workflows. The frontend is available as either a lightweight HTML/CSS/JavaScript interface or a production-grade Next.js application with Tailwind CSS. The default retrieval backend uses Tavily's API, though other search providers can be configured [9].
| Detail | Value |
|---|---|
| Founded | Late 2023 |
| Founders | Rotem Weiss (CEO), Assaf Elovic (Co-founder) |
| Headquarters | New York, NY |
| Employees | Approximately 30 (as of 2025) |
| Total funding | $25 million |
| Key investors | Insight Partners, Alpha Wave Global |
| Acquisition | Agreed to be acquired by Nebius (February 2026) for up to $400 million |
| Notable customers | IBM, AWS, MongoDB, Cohere, Groq, JetBrains, Monday.com, LangChain |
| Certifications | SOC 2 |