Tavily AI Search: The Search API Built for LLMs, Not Browsers

17,789 downloads, 224 installs, 24 stars. The Tavily AI Search skill by @bert-builder is the most installed web-search skill on ClawHub by a significant margin. The reason comes down to a fundamental design difference: every search API on the market was built for browsers first. Tavily was built for LLMs first.

The Problem It Solves

When a Claude agent calls a traditional search API, it gets back raw SERP results — titles, URLs, and snippets designed for human scanning. The agent then has to decide what's relevant, fetch the actual pages, parse out the useful text, and synthesize an answer. That's four steps where the agent could go wrong, and four steps burning tokens.

Tavily short-circuits this. A single API call returns ranked results plus an AI-generated answer summary from those results, plus clean extracted content from each page — all formatted for LLM consumption. What used to require a search skill, a scraper, and a summarizer can now happen in one call.

How It Works

The skill wraps the tavily-python SDK and exposes a CLI that takes natural language queries and returns structured JSON. The underlying architecture is straightforward:

Query → Tavily API → Domain Filter → Web Crawl → AI Answer Generation → Structured Response

Two modes determine the depth of this pipeline:

Deep Dive

The AI Answer Layer

The most distinctive feature. When include_answer=true (the default), Tavily synthesizes an answer from its search results before returning them to you. The agent gets both the synthesis and the sources — it can use the answer directly or verify it against the raw results.

# Basic search with AI answer
python scripts/tavily_search.py "What changed in Python 3.12?"
 
# Get output with AI-synthesized answer first, then ranked sources

This is fundamentally different from DuckDuckGo (billyutw/web-search) which returns snippets only, or Brave Search (steipete/brave-search) which returns ranked results only. Neither generates a synthesized answer from the sources.

Domain Filtering

Precise control over which sources the search draws from — either allowlist (include) or denylist (exclude):

# Only search authoritative Python docs
python scripts/tavily_search.py "asyncio best practices" \
  --include-domains python.org docs.python.org realpython.com
 
# Exclude low-signal sources
python scripts/tavily_search.py "React hooks tutorial" \
  --exclude-domains w3schools.com geeksforgeeks.org
 
# Research mode with domain control
python scripts/tavily_search.py "SEC filing requirements 2026" \
  --depth advanced \
  --include-domains sec.gov investor.gov

This is particularly valuable for agents doing compliance research, technical documentation lookup, or competitive intelligence where source quality matters.

News Mode

When a query is about current events, switch the topic to news:

# Breaking news / recent developments
python scripts/tavily_search.py "AI regulation updates" --topic news
 
# Combined: news mode + specific domains
python scripts/tavily_search.py "Fed interest rate decision" \
  --topic news \
  --include-domains reuters.com bloomberg.com

News mode limits results to the last 7 days and prioritizes news sources. Particularly useful for agents that need to distinguish between evergreen content and recent developments.

Raw Content Extraction

Beyond snippets, Tavily can fetch and clean the full page content:

# Get full article text from sources
python scripts/tavily_search.py "how Rust ownership model works" \
  --raw-content \
  --max-results 3

With include_raw_content=true, each result includes the full cleaned HTML of the source page. This gives the agent the same information a human would get from clicking through and reading — without needing a separate scraping tool.

Image Search

For multimodal workflows:

python scripts/tavily_search.py "React component architecture diagram" --images

Returns relevant image URLs alongside text results.

How to Install

clawhub install tavily

Get your API key at tavily.com. The free tier includes 1,000 credits/month (no credit card required). A basic search costs 1 credit; advanced costs 2. Pay-as-you-go is $0.008/credit beyond the free tier.

export TAVILY_API_KEY="tvly-your-key-here"

Comparing Search Skills on ClawHub

Bottom line: If you want raw search results with a strong independent index and privacy guarantees, use brave-search. If you want zero-setup search for prototyping, use billyutw/web-search. If you want AI-synthesized answers, domain-controlled sources, and results explicitly formatted for LLM consumption — use Tavily.

Practical Tips

Use basic by default, upgrade selectively. Most queries don't need advanced. Reserve advanced mode for research tasks where completeness matters more than speed.

Domain filtering is an underused superpower. For any domain-specific agent (legal, medical, technical), hardcoding trusted source lists turns Tavily from a general search tool into a precision retrieval system.

News mode for agent loops. If your agent runs periodically (hourly, daily) and needs to catch recent changes, --topic news ensures it's not recycling stale results from months-old pages.

Combine with --no-answer for retrieval pipelines. For RAG workflows where you want to embed and chunk source documents yourself, skip the AI answer layer with --no-answer and use --raw-content to get clean text.

Considerations

• API cost. Tavily is not free at production scale. Each search consumes credits. Check current pricing at tavily.com before deploying to high-volume workflows.
• No independent index. Unlike Brave Search, Tavily's index is backed by partner sources. For queries where source independence matters, brave-search is the better choice.
• 5–10s latency in advanced mode. Not suitable for latency-sensitive workflows. Plan agent pipelines around the worst-case timing.
• AI answer quality varies. The synthesized answer is only as good as the source quality. Domain filtering helps, but verify the answer against raw results for high-stakes queries.

The Bigger Picture

Tavily is one of the first search APIs designed with the assumption that its primary consumer is an AI agent, not a human. The features that define it — answer synthesis, raw content extraction, domain filtering — all exist because agents have different needs than browsers. They don't scan result lists; they need synthesized, consumable data.

As AI agents become the primary interface between humans and information retrieval, the gap between "browser-optimized" and "agent-optimized" search will widen. Tavily is an early example of what that second category looks like.

View the skill on ClawHub: tavily