Tavily

💰 Pricing & Plans

🔧 API & Developer Experience

• •Five endpoints, one credit system: Search, Extract, Crawl, Map, and async Research Task — all share one credit pool so multi-step research workflows never juggle separate vendor SKUs.
• •First-class Python & JavaScript SDKs: `pip install tavily-python` and `npm i @tavily/core` both ship with sync/async clients; LangChain, LlamaIndex, AutoGen, and CrewAI ship Tavily as the default web search tool out of the box.
• •MCP server for coding agents: Tavily publishes a Model Context Protocol server, so Claude Code, Cursor, and Codex get zero-integration web search inside the editor — no glue code, no API key plumbing in agent prompts.
• •Credit-based, predictable billing: 1 Search / Extract / Crawl call = 1 credit (advanced options cost 2-5 credits). PAYG = $0.008/credit; Project = $30+ with 4,000 credits. Hard ceiling, no surprise per-token meters.
• •Real-time web, AI-extracted answers: Search returns both raw ranked URLs and a generated AI summary/answer, so RAG pipelines get cited evidence in one round-trip instead of crawling + LLM-summarizing separately.
• •Free tier with no credit card: Researcher plan = 1,000 credits/month, zero friction — enough to prototype a multi-agent research workflow end-to-end before committing to a paid plan.
• •Native async research agent: Research Task endpoint runs multi-step web research server-side and polls a job ID, offloading long research from the client and avoiding timeout-prone synchronous chains.

🔍 Search-as-a-Tool for Agent Frameworks

Tavily's defining bet is treating web search as a turnkey tool layer rather than a raw retrieval API. Where traditional search engines return ten blue links, Tavily's `/search` endpoint returns both ranked URLs and a synthesized AI answer with inline citations — a single round-trip that gives RAG pipelines usable evidence instead of forcing them to crawl, dedupe, and summarize separately. The company ships official LangChain, LlamaIndex, AutoGen, and CrewAI integrations, so any agent built on these frameworks can call Tavily without custom glue code. Five endpoints (Search, Extract, Crawl, Map, and the async Research Task) share one credit pool and one billing identity, making cost forecasting simple: a research agent that does search-then-extract on N pages is just N+M credits, not three separate vendor invoices. For teams building multi-step research agents on tight budgets, that single-vendor simplicity is the core reason to standardize on Tavily over stitching together Brave / Serper / Firecrawl.

🌐 China Access & Latency

Tavily's API endpoints are hosted in the US (api.tavily.com), so mainland access requires a stable proxy or VPN in production. Latency from a Shanghai or Beijing line via a clean proxy is typically 250-450ms round-trip for Search, dominated by the cross-border link rather than Tavily's processing. The Researcher (free) tier works fine behind most commercial proxies, but for production agents handling high RPS the more important question is whether your agent runtime itself can reach the endpoint — Tavily does not ship a mainland mirror or an ICP-registered edge. Teams that need to stay entirely within China's network boundary should pair Tavily with a self-hosted search alternative (Bocha, Tongyi, or a self-run SearXNG) for the data-residency-sensitive steps, and use Tavily only for the queries that genuinely need global coverage. Documentation, signup, and the dashboard are all reachable from China without VPN, so onboarding and configuration are friction-free even when the API itself needs a proxy.