Web Search Skills for AI Agents: Tavily, Brave, Exa, SearXNG & More
Every AI agent that touches the real world needs one thing before reasoning: fresh information. Web search skills are the retrieval layer — the pipe between the open internet and your agent's context window.
ClawHub currently hosts web search skills across 8 major technical approaches — from managed APIs to self-hosted privacy stacks.
Note: Install and download figures in text descriptions reflect stats at the time of writing and may be outdated. All skill tables are live — they fetch current data from the ClawHub database on every page load. Treat table values as authoritative.
Data Overview
| Metric | Value |
|---|---|
| Total web search skills (approx.) | 342 auto-classified |
| Top by downloads | tavily-search ( downloads) |
| Top by installs | tavily-search ( installs) |
| Downloads ≥ 10,000 | 8 skills |
一、Tavily — The Dominant Search API
Tavily has become the de facto standard for LLM-native web search. Its API returns concise, relevance-ranked results optimized for AI consumption rather than human browsing — no HTML noise, no pagination. tavily-search alone has over 161,000 downloads, making it the most-downloaded web search skill on the platform by a wide margin.
The sheer number of Tavily variants (tavily-search, openclaw-tavily-search, tavily, liang-tavily-search...) reflects a community pattern: developers fork the reference implementation and customize the system prompt, endpoint, or API key management. The original and the OpenClaw-official fork together account for over 200,000 downloads.
二、Brave & Exa — Privacy-First and Neural Search
Brave Search (33,601 downloads, 413 installs) is the second-largest non-Tavily search skill. It hits Brave's independent index — not a Google reskin — making it genuinely alternative data. Useful when you want search results that aren't shaped by Google's ranking signals.
Exa (formerly Metaphor) takes a different approach: neural/semantic search rather than keyword matching. web-search-exa (21k downloads) and exa-web-search-free (12.4k) are particularly popular for research workflows where the query is conceptual ("papers about X") rather than keyword-precise.
三、Multi-Engine & Hybrid Search
Single-source search has a failure mode: when one API is down or rate-limited, the whole agent stalls. Multi-engine skills route across multiple backends with auto-fallback.
web-search-plus (13.2k downloads) uses intelligent auto-routing across Tavily, Serper, and SearXNG. multi-search-engine (12k downloads) covers 74 installs — unusually high ratio, suggesting it's in active daily use. super-websearch-realtime runs parallel queries and merges results.
四、Serper & SerpAPI — Google SERP Access
Serper and SerpAPI both wrap Google Search results via paid APIs. Use these when you specifically need Google's ranking, Google News, or Google Shopping data — not available through Tavily or Brave.
serper (3k downloads) returns structured SERP data with full page extraction. serpapi-mcp (1.2k) integrates as an MCP server for direct use inside Claude.
五、Firecrawl — Web Crawling + Search
Firecrawl is different from the others: it's a crawling and scraping API first, search second. You point it at a URL or a domain, and it returns clean markdown — ideal for extracting full page content that a search snippet doesn't capture.
firecrawl-search (9.2k downloads) combines Firecrawl's search endpoint with its content extraction. firecrawler (2.8k) focuses on single-page scraping with AI-ready output.
六、SearXNG — Self-Hosted Privacy Search
SearXNG is an open-source metasearch engine you deploy yourself — it aggregates results from 70+ sources without sending your queries to any third-party API. For air-gapped environments, enterprise privacy requirements, or developers who don't want to pay per-query, it's the only real option.
The high number of SearXNG variants on ClawHub (30+) reflects that every self-hosted deployment has a different endpoint URL — hence each user publishes their own fork. They're functionally identical; pick any one and substitute your instance URL.
七、Jina Reader — URL to Markdown
Jina Reader converts any URL into clean, LLM-readable markdown. It's not a search engine — it's the last mile: once you know which page you want, Jina fetches it and strips the HTML noise. Essential complement to any search skill.
八、Chinese Search Engines
For Chinese-language research, Baidu, Zhipu, and Minimax provide access to content that Western search APIs systematically miss. cn-web-search (4.9k downloads, 52 installs) is the most widely used Chinese search skill.
How to Choose
| I need... | Recommended skill |
|---|---|
| Best overall LLM-native search | tavily-search |
| Google's actual ranking | serper |
| Semantic / conceptual queries | web-search-exa |
| Privacy-first, no third-party API | searxng + self-hosted instance |
| Fetch full page content | firecrawl-search or jina-reader |
| Multi-source with auto-fallback | web-search-plus |
| Chinese-language content | cn-web-search |
| Free tier, no API key | exa-web-search-free |
Last Words
Tavily's dominance is structural, not just popularity. 161k downloads vs. Brave's 33k isn't a preference gap — Tavily is the only search API purpose-built for LLM consumption. Its results are pre-summarized, citation-ready, and consistently sized. For agent workflows, this matters more than raw index size.
SearXNG has 30+ variants but zero centralized install count. This is the telemetry problem made visible: every private SearXNG deployment gets its own skill with its own slug. Collectively they represent a large self-hosted search community that install counts entirely miss.
The Exa vs. Tavily choice is a query-type choice. Tavily excels at factual, news, and recent-events queries. Exa excels at semantic discovery ("find papers like this one," "companies similar to X"). Both belong in a well-equipped agent.
Firecrawl fills the gap search APIs leave. Search returns snippets; Firecrawl returns full documents. For any agent that needs to read a source, not just find it, Firecrawl or Jina Reader is the necessary second step.
Data source: ClawHub platform download and install stats as of April 16, 2026. Browse more skills at clawhub-skills.com.