Tavily

Give AI agents live web access through Tavily. Agents run ranked web searches, extract clean markdown from batches of URLs, crawl and map entire sites, launch deep-research tasks that return cited reports, and monitor API credit usage.

7 actions API key available

Tavily website ↗ Platform Documentation ↗ Capability data verified 2026-07-22

The quarterly knowledge-base refresh comes due for a partner's documentation site

Map Website traverses the site and returns the full URL inventory without extracting anything

The agent sizes the job from the map: page count, projected credits at crawl rates, and the remaining budget from Get API Usage

The owner approves the crawl scope, trimming the paths that do not belong in the knowledge base

Crawl Website extracts the approved sections as clean markdown, following internal links to the configured depth

The content lands in the knowledge base with source URLs preserved per chunk

A completion summary with key links posts to [Slack](/integrations/slack)

What This Integration Enables

Tavily is web access built for agent pipelines, and its most operationally honest feature is that everything has a price in credits: search, extraction, crawling, mapping, and research each meter separately, and Get API Usage exposes the meter to the flow itself. That turns web access from an unbounded cost into a governed resource. FlowRunner agents run ranked searches with domain and date filters, extract clean markdown from batches of URLs, map and crawl entire sites, and launch deep-research tasks that come back as cited reports or structured JSON. The [human-in-the-loop](/concepts/human-in-the-loop) discipline applies where the credits concentrate: big crawls and deep research get scoped by a person before they spend. - Run AI-optimized searches across general, news, and finance topics with domain and date filters - Extract clean markdown from up to 20 URLs per call, with failures reported per URL instead of failing the batch - Map site structures cheaply before committing to full crawls - Launch deep-research tasks with Mini speed or Pro depth, returning cited reports or schema-shaped JSON

Without FlowRunner

Agents answer from training data Questions about live facts get confident answers frozen at the model's cutoff

Scraping is a side project Brittle parsing scripts break on every markup change and nobody owns them

Web costs surface on the invoice Nobody knows which flow spent the credits until the bill arrives

With FlowRunner

Results arrive agent-ready Ranked, scored, snippeted, and already markdown, with no parsing stack to maintain

Whole sites become corpora Map, then crawl, then extract turns a documentation site into ingested knowledge

Every operation has a meter Per-endpoint usage is queryable mid-flow, so expensive calls gate on real budget

Use Case Scenarios

A news watch with a paper trail

The team tracks a regulatory topic where being a day late matters. On a schedule, the agent runs Search on the news topic with a relative date range and the trusted-source domain list, then writes each new finding with its relevance score and source link into a research database in [Notion](/integrations/notion). Items crossing a materiality threshold post immediately to [Slack](/integrations/slack). The archive builds itself, and every claim in it traces to the page it came from.

A documentation site turned searchable index

Before ingesting a vendor's documentation, the agent runs Map Website with path filters to discover the structure, then feeds the relevant URLs to Extract Content in batches of 20, logging each page's markdown and summary to [Airtable](/integrations/airtable). URLs that fail extraction arrive in failed_results with reasons, so the agent retries or flags them instead of silently shipping gaps. The result is a content index a support agent can query, built without writing a single parser.

Deep research sized to the question

Not every question deserves the same spend. For a quick landscape check, the agent calls Start Research with the Mini model and gets a cited report in under a minute. For a build-versus-buy decision, it uses Pro with a JSON Schema so the output lands as structured fields in the evaluation doc, published through [Google Docs](/integrations/google-docs). Get Research Results polls until completion, and the reviewer who owns the decision walks the cited sources behind the load-bearing claims before the report circulates as fact.

Human-in-Loop Highlight

Crawl Website is where credits concentrate: it bills per pages extracted, natural-language crawl instructions double the per-page rate, and a site whose depth nobody checked can turn a routine ingestion into a four-figure page count. So the flow never crawls blind. Map Website runs first at the cheaper mapping rate to produce the actual inventory, the agent lays the page count and projected credits beside the account's remaining budget from Get API Usage, and a person approves the scope, or trims it, before extraction begins. Start Research on the Pro model gets the same treatment on recurring schedules: depth is a spend decision, and spend decisions have owners.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

7 actions

Search

Search Runs an AI-optimized web search returning ranked results with relevance scores and snippets. Supports general, news, and finance topics, optional LLM-generated answers, raw page content, image results, include and exclude lists up to 300 and 150 domains, date filtering, country boosting, and exact phrase matching. Auto Parameters lets Tavily tune settings per query, with explicit values taking precedence. Up to 20 results per call.

Extraction

Extract Content Extracts up to 20 URLs per call into clean markdown or text, with Advanced depth capturing tables and embedded content at a higher credit rate. An optional query reranks extracted chunks by relevance. Failed URLs are itemized with errors rather than sinking the batch.

Crawling

Crawl Website Crawls from a base URL through internal links to configured depth and breadth, returning every page's extracted content. Natural-language instructions focus the crawl at double per-page cost, and regex filters constrain paths and domains. Built for ingesting documentation sites and knowledge bases.
Map Website Traverses a site and returns discovered URLs without extracting content, a sitemap-style overview at the cheapest rate. The scouting step that sizes a crawl before it spends.

Research

Start Research Starts an asynchronous deep-research task that plans, searches, and synthesizes a cited report. Mini for speed, Pro for exhaustive depth, Auto to let Tavily choose. Supports domain preferences and blocklists, citation formatting, report length control, and a JSON Schema for structured output.
Get Research Results Polls a research task until completed, then returns the report as markdown or schema-shaped JSON along with the sources used. Mini tasks finish in under a minute; Pro can take several.

Account

Get API Usage Returns credit usage for the key and the account this billing cycle, with per-endpoint breakdowns across search, extract, crawl, map, and research, plus plan and pay-as-you-go limits. The instrument that lets flows gate their own spending.

Frequently Asked Questions

What can FlowRunner do with Tavily?

FlowRunner agents can run Search, Extract Content, and Crawl Website in Tavily, plus 4 more actions.

Does connecting Tavily to FlowRunner require OAuth?

No. Tavily connects to FlowRunner with an API key, no OAuth flow required.

Can Tavily trigger a FlowRunner workflow automatically?

Tavily doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with Tavily

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us