Tavily

  • Tavily — Full Profile (2026-09)

    Tavily is an Alioth-adjacent AI research tool operating with a NYC/remote-first model, founded in 2023 by Rotem Weiss and Assaf Elovic, spun out of the open-source GPT Researcher project. The company completed successive financing rounds including a mid-2024 seed round and a $20M Series A led by Insight Partners and AlphaWave in August 2025, bringing total capital raised to $25M. In February 2026, Tavily was fully acquired by Nebius Group for a base value of $275M, with potential milestone-based uplifts reaching up to $400M. Post-acquisition, the product maintains independent operation under the original Tavily brand within Nebius AI Cloud; it retains full business continuity and is not absorbed or sunset, distinguishing it from sunset-acquired tools such as MonkeyLearn. Tavily offers an official, stable first-party MCP server hosted at mcp.tavily.com/mcp, complemented by an open-source MIT-licensed repository tavily-ai/mcp-server-tavily with over 1.2K GitHub stars. Its MCP tool suite covers core web intelligence capabilities including search, extract, map, crawl, and end-to-end research generation. Ecosystem-wise, Tavily belongs to the stable first-party GA server cluster alongside Klaviyo, Feedly, Slack and Databox. It is clearly differentiated from n8n dual-GA architecture, Google Sheets/Gmail first-party preview stack, Typeform beta-grade MCP, Nanonets Composio-exclusive integration model, and the zero-first-party MCP cohort consisting of Documenso, QuickChart and Semantic Scholar.

    Website

    Field Value
    Official Website https://tavily.com · docs: https://docs.tavily.com · API base: https://api.tavily.com · MCP host: https://mcp.tavily.com/mcp · OSS MCP: github.com/tavily-ai/mcp-server-tavily · SDK: github.com/tavily-ai/tavily-python · github.com/tavily-ai/tavily-js
    Company Tavily (founded 2023, NYC/remote), acquired by Nebius Group Feb 2026 (125M milestones), operates as Tavily-under-Nebius AI Cloud; founders Rotem Weiss + Assaf Elovic; grew from GPT Researcher OSS (27K★); 700K-2M developers, 100-300M req/mo, 180ms p50 search, 99.99% uptime SLA on paid; SOC 2 Type II + GDPR, prompts not used for training, no DPA on free tier; Cloud-only, no on-prem
    Status Active, 5 endpoints (Search/Extract/Map/Crawl/Research), official MCP server GA, Nebius-integrated
    Category ai-productivity-data
    Pipeline Stage none

    Title

    Field Value
    Product Tavily Web Access API for AI Agents (+ official MCP server)
    Descriptive Title Tavily — Agent-Optimized Web Search/Extract/Crawl/Map/Research API: POST api.tavily.com/search with query+search_depth(basic/advanced)+topic(general/news)+include_answer+include_domains → returns LLM-ready JSON (clean content chunks, relevance scores, optional synthesized answer+citations, no HTML/ad noise); Extract pulls structured text from URLs (1-2 cr/5 ok), Map returns sitemap (1-2 cr/10 pages), Crawl = map+extract (3-5 cr/10 pages), Research runs multi-step agentic loop (4-250 cr); credit-metered not per-seat​ (Researcher 1K/mo free → Project 100/15K → Startup 500/100K → PAYG 25M funding, acquired-kept under Nebius Feb 2026, joins Klaviyo/Databox/Feedly/Slack first-party-GA-MCP cluster
    One-line Positioning Agent calls Tavily → gets pre-cleaned markdown chunks + scores + optional cited answer instead of raw SERP HTML → Claude mounts mcp.tavily.com/mcp with tvly- key, no HTML parsing code; billing is credit-per-endpoint-depth (reverse Motion per-seat / reverse Qualtrics per-interaction / reverse Databox AI-credit-but-similar-shape / reverse Gmail 80M-unit / reverse QuickChart image-count); ownership Nebius NASDAQ:NEBIUS public-traded acquirer but Tavily card stays acquired-kept-brand-retained (not platform-primitive-of-GOOG/META, not vc-independent-anymore).

    Features

    Field Value
    Search POST /search query+search_depth(basic=1cr/advanced=2cr)+topic(general/news)+time_range+start/end_date+max_results(≤20)+include_answer(bool)+include_raw_content+include_images+include_domains/exclude_domains+country+chunks_per_source+auto_parameters(off recommended); returns results[] with title/url/content/snippet/score, optional answer, 180ms p50
    Extract POST /extract urls[]+extract_depth(basic/advanced) → clean markdown/text per URL, 1cr/5 ok (basic) or 2cr/5 ok (advanced), failed URLs not charged
    Map POST /map url+instruction(optional) → sitemap URLs, 1cr/10 pages or 2cr/10 with NL instruction
    Crawl POST /crawl url+max_depth+limit+extract_depth → traverse site, cost = map cost + extract cost (e.g. 10 pages basic = 1+2 = 3cr)
    Research POST /research query+model(mini/pro) → multi-step search→read→gap→search→synthesize, 4-110cr (mini) or 15-250cr (pro) dynamic
    MCP (first-party GA) Official​ https://mcp.tavily.com/mcp?tavilyApiKey=... (remote stdio/streamable-http) + OSS tavily-ai/mcp-server-tavily MIT ~1.2K★; tools: tavily_search / tavily_extract / tavily_map / tavily_crawl / tavily_research; Claude Desktop/Cursor/Cline/Windsurf compatible; key-in-URL or OAuth; DEFAULT_PARAMETERS header tunable
    SDK / Framework tavily-python (sync+async), @tavily/core (JS), LangChain TavilySearchResults native, LlamaIndex native, CrewAI native, OpenAI/Anthropic/Groq tool-use wrappers, Databricks MCP Marketplace, JetBrains in-IDE
    Safety PII detection (blocks email/phone in query), prompt-injection filtering, content-validation layer, SOC 2 Type II, GDPR, no-training-on-prompts, X-Project-ID header for multi-tenant billing split

    Pricing (2026-09, credit-metered per endpoint×depth, no seats, free 1K/mo permanent)

    Field Value
    Researcher (Free) $0​ — 1,000 API credits/mo, no card, all endpoints, 100 RPM dev limit, community email support
    Pay As You Go $0.008/credit​ — overflow beyond plan pool, cancel anytime
    Project $0.0075/cr), higher rate limits
    Bootstrap $0.0067/cr)
    Startup $0.0058/cr)
    Growth $0.005/cr), 1,000 RPM prod
    Enterprise Custom — custom credits, rate limits, SLA, DPA, on-prem/cloud-VPC, SOC 2 reports, students free
    Credit costs Search basic 1 / advanced 2; Extract 1-2 per 5 ok URLs; Map 1-2 per 10 pages; Crawl = map+extract; Research mini 4-110 / pro 15-250 per call
    Type usage-credit-metered-no-seat; reverse Motion per-seat, reverse Qualtrics per-interaction, reverse Nanonets block-run, closest to Databox AI-credit but credit=endpoint-depth not AI-tokens, reverse Gmail 80M-unit, reverse QuickChart image-count, reverse Semantic Scholar 1-RPS
    Gotcha auto_parameters=true silently upgrades basic→advanced (2cr); unused credits expire monthly; Research pro can burn 250cr/call — loop guard needed

    Reviews

    Field Value
    G2 4.8/5 on 2 reviews​ (AI Agent Index mirror, unclaimed profile, not robust sample) → print G2 4.8/~2 unclaimed-thin (do NOT treat as 4.8-solid like Klaviyo 4.4/~580); dev-tool G2 pattern weak
    Capterra None​ — no standalone Capterra page → no-standalone-capterra-profile
    Trustpilot No B2C page​ (B2B dev infra) → no-rated-consumer-profile excluded (aligns FFmpeg/QuickChart/Semantic Scholar)
    GitHub / Dev sentiment MCP server 1.2K★, tavily-python active, Saashub/Dev.to 4.5-ish directional; AIMultiple agentic benchmark 13.67/20 (behind Brave 14.89, ahead of Perplexity 12.96), latency 998ms (sub-1s), quality 3.77/5; praise “cleanest LLM-ready output, best free tier, LangChain-native”; gripes credit burn on advanced/research, aggregator-not-own-index (upstream Bing price risk), keyword-retrieval weaker than Exa neural, no DPA on free, acquired-by-Nebius roadmap uncertainty
    Editorial AI Agent Index 4.7/5 composite (Autonomy/Integrations/Pricing 5, Evidence/Setup 4); ToolRadar-style 4.6 norm; “Stripe for agent search” narrative
    Praise Pre-cleaned chunks+scores+citations, sub-1s p50, free 1K/mo no card, official MCP GA (not community-only), LangChain/LlamaIndex zero-config, Research endpoint replaces multi-step agent loop, Nebius acquisition = infra durability
    Gripes Credit opacity (auto_parameters, Research 250cr), not own index (aggregator), Exa beats on semantic/obscure-source, Brave beats on latency+fresh-news, no DPA free-tier, PII block can false-positive, acquired status = pricing could shift under Nebius

    Review Sources Breakdown

    Field Value
    G2.com 4.8/5 ~2 reviews unclaimed → G2-thin-unclaimed-2-review-not-solid (flag vs Klaviyo/Capterra-rated SaaS)
    Capterra none → no-standalone-capterra-profile
    Trustpilot none → no-rated-consumer-profile excluded
    AIMultiple benchmark 13.67/20 agent score, 998ms, 3.77/5 quality — directional not star
    GitHub 1.2K MCP + dev blogs directional dev-sentiment
    AI Agent Index 4.7 composite not G2-grade
    Conclusion​ Prints “G2 4.8/~2 unclaimed-thin, Capterra none, TP none, AIMultiple 13.67/20, GitHub MCP 1.2K, dev consensus LLM-ready-best-free-tier-acquisition-durability, credit-burn-gripe” — aligns QuickChart/Semantic Scholar no-G2-solid rule but Tavily has 2-review G2 trace so cite it thin not zero.

    Access

    Field Value
    REST POST api.tavily.com/search
    MCP First-party GA​ — remote https://mcp.tavily.com/mcp?tavilyApiKey= (stdio/streamable-http) or OSS npx -y mcp-server-tavily with TAVILY_API_KEY env; tools search/extract/map/crawl/research; Claude/Cursor/Cline/Windsurf/ChatGPT-agents compatible
    SDKs tavily-python (async/sync), @tavily/core, community Go/Ruby/PHP; llms.txt docs index for agents
    Frameworks LangChain TavilySearchResults, LlamaIndex TavilyToolSpec, CrewAI TavilySearchTool, OpenAI Assistants function wrapper, Anthropic tool-use snippet, Databricks Marketplace MCP, JetBrains plugin
    Login app.tavily.com email/Google → API Keys page → copy tvly- key
    Self-host None​ — Nebius Cloud only, no on-prem/Docker (contrast QuickChart/Documenso self-host)

    Best For

    Field Value
    Audience 1 LLM-agent / RAG builders needing web results as clean chunks+scores+citations, not SERP HTML
    Audience 2 LangChain/LlamaIndex/CrewAI projects wanting zero-config TavilySearchResults tool
    Audience 3 Claude/Cursor agents mounting official MCP for search/extract/crawl without writing HTTP glue
    Audience 4 Research-agent loops using /research to replace multi-step search-synthesize code (cost-aware)
    Audience 5 Startups wanting 1,000 free searches/mo no card before $30 Project
    Audience 6 Enterprises needing SOC 2 + GDPR + no-training-on-prompts for agent web access (paid tiers)

    Not For

    Field Value
    Exclude 1 Teams needing own web index / independent crawl (Tavily aggregates upstream Bing+curation, not Brave/Exa-style own-index) — use Brave/Exa
    Exclude 2 Sub-200ms high-frequency algos (Tavily p50 180ms but p95 1s+, advanced 2-3s) — use Brave
    Exclude 3 Semantic/obscure-source discovery where keyword retrieval fails — Exa neural wins
    Exclude 4 Buyers refusing credit-metered volatility (Research pro = 250cr/call) — use Serper $0.001/query raw SERP
    Exclude 5 On-prem/air-gap agent search — Cloud-only, no self-host
    Exclude 6 Free-tier users needing DPA/SLA — only on Enterprise
    Exclude 7 Non-developer/no-code-only teams without API integration capacity

    Competitors

    Field Value
    Exa (Metaphor) Neural embeddings index, semantic link-prediction, ~$1.50/1K searches, better obscure-source/RAG-embedding fit, official MCP, no answer-synthesis legacy (now has Exa Answers)
    Brave Search API Independent 30B-page index, $5/1K AI-grounding tier, 669ms fastest, best fresh-news, no LLM-ready chunking (you parse), official MCP
    Perplexity Sonar Answer-first API, cited markdown answer not raw results, $5/1K-ish, not endpoint-granular like Tavily
    Firecrawl Crawl+extract focused (full-page render), not search-ranked, €14-71/mo tiers, complementary (Tavily Crawl overlaps)
    Serper / SerpAPI Raw Google SERP JSON, Serper $0.001/query cheapest, no cleaning/extract, devs build parsing themselves
    Linkup / Parallel / You.com API Agent-search mid-field, smaller adoption than Tavily-Exa-Brave triad

     

Do Not Sell or Share My Personal Information Cookie Settings