Tavily

Tavily is a search and content extraction API built specifically for AI agents and large language model applications. Founded in late 2023 by Rotem Weiss and Assaf Elovic in New York, the company grew out of GPT Researcher, an open-source project Elovic launched in July 2023 to give LLMs real-time web access before ChatGPT had native browsing. The commercial product solves a concrete problem: when an LLM-powered agent needs current information, it needs search results formatted for a context window, not an HTML page with cookie banners and navigation menus mixed in. Tavily handles that formatting layer, returning ranked snippets, relevance scores, and source URLs ready to pipe directly into a prompt.

The API offers four main endpoints: Search (keyword and AI-ranked real-time results), Extract (structured content from specific URLs), Map (crawl and index a domain), and Research (an autonomous multi-step research workflow). Basic Search costs 1 credit per request; Advanced Search costs 2. The free Researcher tier includes 1,000 credits per month. Paid tiers start at $30/month for 4,000 credits. Tavily ships official integrations with LangChain (the langchain-tavily PyPI package), LlamaIndex, the OpenAI Agents SDK, IBM WatsonX, AWS, Databricks MCP Marketplace, and JetBrains IDEs. As of February 2026, the company was acquired by Nebius (NASDAQ: NBIS) for an initial $275 million, with the platform continuing to operate under the Tavily brand.

What Tavily actually does in April 2026

Tavily is not a general web search engine with a UI. It is an API-first infrastructure layer that connects agent loops to the live internet. A typical call sends a query string and optional parameters (topic, time range, domain filter, max results) and receives back a JSON object with ranked results, each containing a title, URL, snippet, and relevance score. For advanced use cases, the Extract endpoint fetches and parses specific pages, stripping boilerplate to return clean markdown. The Map endpoint crawls an entire domain and returns a list of indexed URLs, useful for building knowledge bases from a specific site.

The Research endpoint is the most compute-intensive offering. It accepts a natural language research question and autonomously performs multiple searches, synthesizes findings, and returns a structured report with citations. Credit cost varies significantly: 15-250 credits per request on the Pro model, 4-110 credits on the Mini model. This means a complex research query on pay-as-you-go pricing ($0.008/credit) can cost anywhere from $0.12 to $2.00 per call, which surprises developers who assumed flat-rate pricing across all endpoints.

Production-grade features include a 99.99% uptime SLA, 180ms p50 latency on standard search queries, PII filtering, prompt injection blocking, and malicious source filtering. The platform claims 100 million+ monthly requests and a developer community exceeding one million users. Customers cited in the Nebius acquisition announcement include IBM, Cohere, Groq, MongoDB, and LangChain.

Where Tavily sits versus Exa and SerpAPI

Three APIs dominate the agentic search conversation: Tavily, Exa, and SerpAPI. They work in fundamentally different ways.

Tavily vs. Exa: Exa (formerly Metaphor) maintains a proprietary embeddings-based index. Queries are matched by semantic vector similarity rather than keyword ranking. When a developer asks for "papers arguing that transformer attention is inefficient," Exa searches meaning; Tavily searches text. In practice, Exa's semantic index results in faster responses: average 1.18 seconds versus Tavily's 1.88 seconds in 2025 benchmarks from dev.to's SERP comparison, and p95 latency of 1.4-1.7 seconds versus Tavily's 3.8-4.5 seconds in Fortune 100 enterprise evaluations. Exa also edges Tavily on SimpleQA accuracy: 94.9% versus 93.3%. The tradeoff is freshness. Exa's crawled index can lag breaking news by hours; Tavily's real-time retrieval pulls from live sources, making it the stronger choice for time-sensitive queries. Exa pricing: $5 per 1,000 standard queries, jumping to $25/1,000 for 26-100 results per query.

Tavily vs. SerpAPI: SerpAPI does not maintain its own index at all. It scrapes live results from Google, Bing, Yahoo, Amazon, and 20+ other engines on demand, returning structured JSON that mirrors the real SERP layout including ads, sitelinks, and knowledge panels. This is excellent for SEO monitoring, competitive intelligence, and any use case that needs to know what Google actually returns. It is a poor fit for LLM pipelines: the output is not formatted for context windows, requires downstream parsing, and includes substantial noise. Pricing reflects the different market: $15 per 1,000 searches on the $75/month base plan, making it 3x more expensive than Tavily at comparable volume. Free tier is 250 searches per month versus Tavily's 1,000.

"Tavily has issue with JS-rendered pages. It seems to be doing it offline and then caching it. It's flaky." - LiranYoffe, GitHub community discussion #167015, 2025

"Agentic search is a multi-billion-dollar opportunity..poised to grow exponentially as enterprises deploy autonomous AI systems." - Rotem Weiss, Tavily CEO, Nebius acquisition announcement, February 10, 2026

What the daily integration reality looks like

Tavily's most common integration pattern is as a tool inside an LLM agent loop. In LangChain, this is two lines: install langchain-tavily, initialize TavilySearch as a tool, and pass it to an agent executor. The tool appears in the agent's available actions; the model calls it when it determines a web lookup is needed. The same pattern works in LlamaIndex, Llama Stack, and the OpenAI Agents SDK. For teams not using an orchestration framework, the REST API is straightforward: POST to https://api.tavily.com/search with a JSON body containing your API key, query, and optional filters.

In practice, developers encounter a few recurring integration issues. First, link quality: Tavily pulls from cached and indexed sources without real-time revalidation. Returned URLs may resolve to 404 pages, especially for fast-changing or low-authority sites. Multiple GitHub community threads document this, and the practical fix is adding a validation layer to check HTTP status before processing. Second, JavaScript-heavy pages: Tavily's extraction pipeline appears to render JS offline and cache the result rather than rendering live. Pages that depend on client-side JavaScript for their main content may return incomplete or stale markup. Third, the free tier's 100 requests-per-minute rate limit is a constraint for developers stress-testing agent workflows. The Production environment (paid plans) raises this to 1,000 RPM.

The practical upside is the LLM-ready output format. Developers in independent tests consistently note that Tavily's ranked snippet format saves significant development time versus building custom scrapers that clean and structure web content.

Who Tavily is built for

Tavily fits best for developers building LLM-powered tools that require real-time web grounding. If the core need is preventing hallucinations by giving a model access to current facts, Tavily is the path of least resistance: the integrations are mature, the documentation is maintained, and the output format is optimized for prompt injection. Teams already using LangChain or LlamaIndex get the most leverage because the integrations abstract away the API call entirely.

Enterprise teams with data security requirements benefit from Tavily's SOC 2 certification, PII filtering, and prompt injection blocking. The platform's customer list (IBM, AWS, Cohere, Groq) signals that it meets enterprise security review requirements that many lightweight scraping tools do not.

The free tier (1,000 credits/month) is practical for prototyping: a developer can run roughly 1,000 basic searches per month at no cost. For production agents processing meaningful query volume, the Project plan ($30/month, 4,000 credits) or Bootstrap plan ($100/month, 15,000 credits) cover most small-to-medium deployments.

What Tavily is not

Tavily is not a semantic search engine. If the use case requires finding conceptually similar content (research papers, historical analogies, people with similar professional histories), Exa's embeddings-based index outperforms Tavily on both recall and latency for those queries.

Tavily is not an SERP intelligence tool. If the goal is tracking what Google returns for a given query, monitoring ad placements, or analyzing SERP features like knowledge panels, SerpAPI's live scraping of actual search engine results is the correct tool.

Tavily is not a JavaScript rendering service. Pipelines that need reliable extraction of content from single-page applications or JS-heavy sites should expect inconsistent results and build fallback logic. Dedicated crawling services handle this more reliably.

Teams running very high query volumes (500K+ requests per month) will find Tavily's credit pricing expensive relative to building a hybrid pipeline with a cheaper commodity search layer and selective Tavily calls for queries requiring AI ranking. Finally, developers evaluating long-term vendor lock-in should factor in the February 2026 Nebius acquisition: pricing tiers and API stability are expected to remain consistent, but the roadmap is now set by a larger infrastructure company rather than the founding team's original vision.