Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.tavily.com/llms.txt

Use this file to discover all available pages before exploring further.

Tavily is the web layer for AI agents. Use Tavily when an agent needs live web search, page extraction, site crawling, site mapping, or cited research. This page answers three questions, in order:
  1. Which Tavily docs should an agent fetch?
  2. How should I connect to Tavily?
  3. Which Tavily capability should I use?

​
1. Agent-readable docs

Every Tavily docs page is also served as clean Markdown — append .md to any docs URL. The whole site is indexed in llms.txt, with the full text in llms-full.txt.

​
2. Choose how to connect

Choose based on what Tavily must do, not only where you happen to be running.

SDK / API

Build Tavily into an application, backend, or agent runtime that calls it in production. Get a free API key at app.tavily.com.

MCP

Make Tavily available across a team or org — governed, shared, and reusable, including standardizing it for everyone’s Cursor and Claude Code. Connect the remote server with OAuth or an API key.

CLI + Skills

Add Tavily to your own local coding or research session — for a single developer, not a team-wide rollout. Install the CLI and run tvly login.
Quick rules
  • Production app or runtime → SDK / API.
  • Shared across a team or org — even in Cursor or Claude Code → MCP.
  • Just your own local session → CLI + Skills.
  • Terminal access alone doesn’t make CLI the right fit — match the choice to where Tavily ultimately needs to run.
Developer resource — Agent Toolkit. Agent Toolkit is for developers building production research agents, not for agents configuring Tavily autonomously. Reach for it when you need deeper research flows, retrieval orchestration, deduplication, summarization, or structured outputs.

​
No account? Connect without a key

For autonomous agents that can’t manage credentials, Tavily offers two no-key paths.

Keyless

Call Search and Extract with zero setup — send the X-Tavily-Access-Mode: keyless header, or connect the remote MCP with no key. No account required.

x402 payments

Pay per request for Advanced Search in USDC on Base — no API key, no account, no human in the loop.

​
3. Choose a capability

Lead with Search when sources are unknown; move to the others once you have URLs or a site to work through.

Search

Start here when sources are unknown or you need current web context.

Extract

Pull clean content when you already have the URL, or one from Search.

Map

Discover a site’s structure before crawling it.

Crawl

Read many pages across a site.

Research

Get cited synthesis — a report, comparison, or decision-ready answer.
These favor quality, which is what most agent workflows need. See Best Practices for Search for the full reference.
  • Prefer search_depth="advanced" for source discovery, comparisons, and high-confidence answers; use "basic" for quick lookups.
  • For latency-sensitive use cases, fast and ultra-fast trade some relevance for lower latency.
  • Add chunks_per_source=3 with advanced Search for stronger evidence per source (chunks require advanced depth).
  • Use max_results=5 for focused answers, 10 for broader research.
  • Use include_domains / exclude_domains when source trust matters.
  • Prefer Search → Extract for grounded answers: Search to find sources, then Extract for full content.
  • Avoid include_answer unless you need a quick answer seed — and still verify against sources.
  • Use Research for cited synthesis: a report, comparison, or decision-ready answer.
A typical agent-grade Search call: 
from tavily import TavilyClient

client = TavilyClient(api_key="tvly-YOUR_API_KEY")
response = client.search(
    "your query",
    search_depth="advanced",
    chunks_per_source=3,
    max_results=5,
)

​
Availability

Some Tavily capabilities, limits, and defaults depend on your account, plan, or enterprise configuration. If a tool or parameter is unavailable, check the relevant endpoint docs and your account settings before retrying a different workflow.