tavily-search: The AI-First Web Search Skill That Replaced Brave for OpenClaw Agents

75,000+ downloads and 341 stars — tavily-search by @arun-8687 is the most-downloaded web search skill on ClawHub. Its rise was partly accidental: when Brave's free search API tier was discontinued, the OpenClaw community needed an alternative fast. Tavily was ready.

What kept it at the top isn't the circumstances of its arrival — it's what it delivers: search results actually designed for AI agents, not humans browsing a webpage.

The Problem With Generic Web Search

When an AI agent uses a standard search API, it gets back what a search engine designed for humans returns: titles, URLs, snippets extracted from metadata, ads mixed in, and verbose structured JSON wrapping it all. The agent has to parse signal from noise. Token budget goes to boilerplate.

Tavily was built with the inverse assumption: the consumer is a language model, not a person. Every design decision — snippet length, result structure, content filtering — optimizes for what an LLM actually needs to reason with.

The result is searches that return less text and more information.

What the Skill Provides

tavily-search wraps two Tavily endpoints into simple Node scripts:

Search

# Basic search — returns 5 results by default
node {baseDir}/scripts/search.mjs "query"
 
# Return more results
node {baseDir}/scripts/search.mjs "query" -n 10
 
# Deep research mode — comprehensive, slower
node {baseDir}/scripts/search.mjs "query" --deep
 
# News-focused search
node {baseDir}/scripts/search.mjs "query" --topic news
 
# Recent news only (last 7 days)
node {baseDir}/scripts/search.mjs "query" --topic news --days 7

Content Extraction

# Extract clean content from a specific URL
node {baseDir}/scripts/extract.mjs "https://example.com/article"

The extract command is particularly useful when search results surface a promising page and the agent needs the full content — not just the snippet.

Search Options Reference

Why Tavily Beats Generic Search for Agents

The community discussion that sparked this skill's adoption put it plainly: Tavily returns "cleaner, more structured, less noise" compared to search APIs that weren't designed with LLM consumers in mind.

Benchmarks back this up. On OpenAI's SimpleQA benchmark, Tavily achieves 93% accuracy — comparable to Perplexity (92%) — with a response time around 998ms. It's not the cheapest option per query, but for tasks requiring accurate, current information, the quality justifies it.

Perplexity's latency problem is significant for agents: a 13-second search blocks the entire session. Tavily's sub-second response means agents can run multiple searches per task without grinding to a halt.

Use Cases Where It Shines

Real-time information: Current events, stock prices, recent releases, breaking news. Use --topic news with --days 3 to get the last 72 hours.

Deep research: --deep triggers Tavily's multi-source aggregation — slower, but pulls from more pages and synthesizes across sources. Appropriate for background research, competitive analysis, or technical deep-dives.

URL-specific extraction: When you know the page you need but want clean text rather than HTML, extract.mjs returns the content in a format LLMs can immediately reason over.

Replacing deprecated search tools: If your current setup relied on Brave Search API's free tier (discontinued) or any other search integration that's become unreliable, this is the drop-in replacement optimized for agent use.

Setup: One Environment Variable

# Get your free API key at https://tavily.com
# Add to your agent settings:

For Claude Code (.claude/settings.json):

{
  "env": {
    "TAVILY_API_KEY": "tvly-YOUR_API_KEY"
  }
}

For OpenClaw (project or user settings):

export TAVILY_API_KEY="tvly-YOUR_API_KEY"

Tavily's free tier is generous for development and light agent usage. The API key is the only setup step — no OAuth, no project configuration, no service account.

How to Install

# Via ClawHub CLI
clawhub install arun-8687/tavily-search

After installation, add your API key to the appropriate settings file. The search scripts are immediately available for the agent to invoke.

Practical Tips

Default to --deep for anything research-grade. The standard mode is fast but shallow. For competitive analysis, technical background research, or anything where accuracy matters more than speed, --deep delivers substantially better results.

Use extract.mjs after finding a promising result. The search endpoint returns snippets. If a result looks valuable, extracting the full page often reveals context that the snippet missed.

Combine search + extract in a single agent action. Search for 5 results, identify the 2 most relevant, extract both, reason over the full content. This pattern gets dramatically better results than relying on snippets alone.

Watch your credit usage on complex queries. Community reports note that multi-step research tasks can consume more credits than expected. Monitor usage in the Tavily dashboard, especially when using --deep at scale.

Considerations

Requires a Tavily API key. Unlike some skills, this one needs a free external account. The signup is fast, but it's a dependency to manage.

Credit consumption varies by query complexity. Simple factual searches are cheap. --deep mode and the extract endpoint use more credits. One community member reported a complex stock summary consuming 147 credits — not the norm, but possible for unusually multi-step queries.

Not open-source web indexing. Tavily's results draw from its own index and selected sources. For queries requiring access to very niche or regional content, coverage may be incomplete.

The Bigger Picture

Web search for AI agents in 2026 isn't a solved problem. The established search giants built their indices and APIs for human browsing patterns. Agents have different needs: structured output, minimal noise, fast latency, and content that can be reasoned over directly.

Tavily was built for this gap. tavily-search packages it into a skill that works immediately — no MCP configuration, no OAuth flow, no multi-step setup. One API key, two scripts, and your agent has reliable web access.

That's why 75,000 OpenClaw users have it installed. Not because there are no alternatives — there are several — but because this one works the way agents actually need web search to work.

View the skill on ClawHub: tavily-search