Tavily
The web-access layer for agents — Search, Extract, Crawl, Map, and Research APIs purpose-built for LLMs, behind one key, with a hosted MCP server.
Tavily packages agent web access as one API: Search tuned for LLM consumption (vendor-claimed 180ms p50), Extract for clean page content, Crawl and Map for site traversal, and a Research endpoint for multi-step investigations — plus SDKs and a hosted MCP server (mcp.tavily.com). Freemium: 1,000 free credits monthly, no card, then pay-as-you-go.
Tavily's framing is exactly the 2026 need: not "a search engine you can call" but the web-access layer for agents — search, extraction, crawling, and multi-step research as one credit pool behind one key, with latency treated as a feature.
Highlights
- Search built for agents — LLM-ready results at basic/advanced depth, with a vendor-claimed 180ms p50 that matters when search sits inside an agent loop.
- Extract, Crawl, Map — clean content from URLs, instruction-guided site traversal, and URL discovery: the ingestion half, included.
- Research endpoint — multi-step investigations (pro/mini tiers) as a single API call, for when one search isn't an answer.
- Hosted MCP server —
mcp.tavily.com/mcp/makes the whole surface a one-liner in Claude Code and friends. - Drop-in ecosystem — Python/JS SDKs and first-class integrations across OpenAI, Anthropic, LangChain, plus marketplace placements (Databricks, JetBrains).
In an AI-assisted workflow
pip install tavily-python # or: npm i @tavily/core
# client = TavilyClient(api_key="tvly-..."); client.search("…", search_depth="advanced")In agent stacks it's typically the web tool: the agentic-RAG searcher, the research agent's eyes, the freshness layer RAG over static corpora lacks.
WARNING
Same caution as every web tool: fetched pages are untrusted input to your model — indirect prompt injection rides in on search results. Treat content as data, and gate any tools that act on it.
Good to know
The company grew out of open-source GPT Researcher and raised ~$25M (a $20M Series A led by Insight Partners, August 2025); in February 2026 Nebius Group agreed to acquire it for $275M (up to ~$400M with milestones), with Tavily continuing under its own brand. It now claims 2M+ developers. SDKs and the MCP server are MIT; the API is the product. Credits aren't 1:1 with calls — budget for advanced/research multipliers. Field positioning against Exa and Firecrawl: Getting Web Data into AI Agents.
Frequently asked questions
- How is Tavily different from Exa and Firecrawl?
- Breadth-in-one-key versus specialist depth. Exa's center is semantic search quality and entity research (Websets); Firecrawl's is industrial-strength scraping and crawling. Tavily bundles credible versions of search + extract + crawl + research behind a single key and credit pool — the convenience pick for agents that need all of it, with speed as its search pitch.
- Is Tavily free to start?
- Yes — the Researcher plan includes 1,000 API credits every month with no credit card; beyond that it's pay-as-you-go per credit or sliding-scale plans. Mind the credit math: advanced search costs 2 credits, and the Research endpoint is dynamic — up to a couple hundred credits for deep runs.
- How do I use Tavily from Claude Code?
- Its hosted MCP server: claude mcp add tavily-remote-mcp --transport http https://mcp.tavily.com/mcp/ (OAuth, or append your tvly- API key as a parameter). It exposes the search/extract/map/crawl tools; a local npm server (tavily-mcp) exists too.
Related
- Getting Web Data into AI Agents: Search & Scraping APIs ComparedThe agent web-data layer — Exa for semantic search, Firecrawl for extraction at scale, Tavily for all-in-one, Jina Reader for zero-setup — and how they compose.
- ExaThe search engine built for AIs — semantic web search, page contents, Websets, and research APIs, plus the ecosystem's most-used search MCP server.
- FirecrawlThe API to search, scrape, and crawl the web for AI — clean Markdown out of any site, LLM-powered extraction, and a first-class MCP server.
- Jina ReaderPrepend r.jina.ai/ to any URL and get LLM-ready markdown — JS rendering, PDFs and Office docs, image captioning, and s.jina.ai for read-the-results search.
- Adding MCP Servers to Claude Code: Local, Remote, and Project-ScopedThe complete claude mcp add reference — stdio vs HTTP transports, local/project/user scopes, .mcp.json with env expansion, OAuth via /mcp, and the gotchas.
- RAG (Retrieval-Augmented Generation)RAG retrieves relevant documents from your own data and injects them into an LLM's prompt at query time, grounding answers in facts the model wasn't trained on.
- Web Research PipelineRun a structured web-research pass on a question: plan the searches, find sources via search APIs, fetch and read the best ones, cross-check claims, and synthesize a cited answer — with source quality and disagreements surfaced honestly. Use for 'research X and tell me what's actually true' tasks that need more than one search and less than a day.