Remote OpenClaw Blog
Best Web Search MCP Servers in 2026: Tavily, Brave, Exa
10 min read ·
Tavily is the best web search MCP server for most AI agents in 2026: it is a search API built specifically for LLMs, its official tavily-mcp server has 2,172 GitHub stars, and the free tier includes 1,000 API credits per month with no credit card. Exa, Brave Search, Firecrawl, and Perplexity round out the top five, and every pick below ships an actively maintained MCP server you can add to Claude Code, OpenClaw, Cursor, or any other MCP client in one command.
This ranking is drawn from the 2,091 servers in the Search category of the Remote OpenClaw MCP directory (13,870 servers indexed in total). It is based on GitHub stars, commit activity, official vendor backing, free-tier generosity, and documented capability as of July 3, 2026, not on paid placement or fabricated benchmark testing.
How We Ranked These Servers
This list ranks the top 8 out of 2,091 search-category servers in our MCP directory using four verifiable signals: GitHub stars, repo maintenance (last push checked July 3, 2026), whether the vendor officially maintains the server, and the free tier on the official pricing page. We did not run a private benchmark, and we say so, because most "we tested 50 servers" posts did not either. Capability notes come from each project's own documentation. If you are new to MCP, start with our MCP servers explainer, and for the cross-category picture see our complete ranked list of the best MCP servers.
1. Tavily: Best Overall Web Search MCP Server
Tavily is a web search API designed specifically for LLMs and AI agents, and its official MCP server (tavily-ai/tavily-mcp, 2,172 stars, MIT) exposes search, extract, map, and crawl tools in one package. Results come back pre-cleaned and relevance-scored for context windows, which is why Tavily is the default search layer in so many agent frameworks. Pricing, per tavily.com/pricing as of July 2026: 1,000 free API credits per month with no credit card, then $0.008 per credit pay as you go.
claude mcp add tavily -e TAVILY_API_KEY=your-key -- npx -y tavily-mcpDirectory page: Tavily MCP.
2. Exa: Best for Deep Research and Semantic Search
Exa is a neural search engine that matches queries by meaning rather than keywords, and exa-labs/exa-mcp-server (4,658 stars, MIT) is the second most-starred server on this list. It adds deep research tools and a code-context search that pulls real usage examples for programming questions, which makes it a strong fit for Claude Code sessions. As of July 2026, Exa's pricing page advertises up to 20,000 free requests per month, then $7 per 1,000 standard searches pay as you go ($12 per 1,000 for deep search).
claude mcp add exa -e EXA_API_KEY=your-key -- npx -y exa-mcp-serverDirectory page: Exa MCP.
3. Brave Search: Best Independent Index
Brave Search runs its own independent web index rather than reselling Google or Bing results, and brave/brave-search-mcp-server (1,260 stars, MIT) is maintained by Brave itself with web, image, video, news, and local search tools. As of July 2026 the Brave Search API gives every account $5 in free monthly credits, which covers 1,000 requests on the $5 per 1,000 Search plan; the newer Answers plan runs $4 per 1,000 requests plus $5 per million tokens. We cover key setup and plan selection in our Brave Search API guide for OpenClaw.
claude mcp add brave-search -e BRAVE_API_KEY=your-key -- npx -y @brave/brave-search-mcp-serverDirectory page: Brave Search MCP.
4. Firecrawl: Best Search Plus Scraping
Firecrawl's official MCP server (firecrawl/firecrawl-mcp-server) is the most-starred server on this list at 6,796 stars, and it bundles web search with the scrape, crawl, map, and extract tools Firecrawl is known for. That combination matters when your agent needs the full cleaned content of result pages, not just snippets. Per firecrawl.dev/pricing as of July 2026: 1,000 free credits per month, then the Hobby plan at $16 per month (billed yearly) for 5,000 credits; search costs 2 credits per 10 results.
claude mcp add firecrawl -e FIRECRAWL_API_KEY=your-key -- npx -y firecrawl-mcpDirectory page: Firecrawl MCP.
5. Perplexity: Best Grounded Answers with Citations
Perplexity's official MCP server (perplexityai/modelcontextprotocol, 2,360 stars) is different from the others here: instead of returning raw result lists, its Sonar API returns synthesized, citation-backed answers from live web research. That makes it the right pick when your agent needs a researched answer rather than links to process. As of July 2026, Sonar pricing is $1 per million input and output tokens plus a request fee starting at $5 per 1,000 requests. See our OpenClaw Perplexity setup guide for a full walkthrough.
claude mcp add perplexity -e PERPLEXITY_API_KEY=your-key -- npx -y server-perplexity-askDirectory page: Perplexity MCP.
6. SerpApi: Best for Google SERP Data
SerpApi returns structured results from Google and other engines, including rankings, shopping, maps, and news verticals, and serpapi/serpapi-mcp is the company's official MCP server (151 stars, MIT, actively pushed as of July 2026). Pick it when the position and structure of Google results is the point, for example rank tracking or SERP analysis agents. Per serpapi.com/pricing as of July 2026: 250 free searches per month, then plans from $25 per month for 1,000 searches. The server runs self-hosted with uv and connects over HTTP:
git clone https://github.com/serpapi/serpapi-mcp.git && cd serpapi-mcp
uv sync && uv run src/server.py
claude mcp add --transport http serpapi http://localhost:8000/YOUR_SERPAPI_KEY/mcpDirectory page: SerpApi MCP.
7. Kagi: Best Quality-Focused Results
Kagi is a paid, ad-free search engine, and kagisearch/kagimcp (432 stars, MIT) is its official MCP server, adding Kagi search plus summarizer tools. The pitch is result quality: no ads and no SEO-spam weighting, which carries over to what your agent reads. The tradeoff is cost and access: the Search API is usage-billed through Kagi's API credit system and Kagi does not publish a flat public rate on its docs page, so check the API portal for current pricing before committing an agent workload to it.
claude mcp add kagi -e KAGI_API_KEY=your-key -- uvx kagimcpDirectory page: Kagi MCP.
8. SearXNG: Best Free Self-Hosted Option
SearXNG is a free, open-source metasearch engine that aggregates results from dozens of engines with no API key and no per-query cost, and ihor-sokoliuk/mcp-searxng (994 stars, MIT) is the most popular MCP bridge to it. You point the server at any SearXNG instance, self-hosted or public, and your agent searches for free with full privacy. The tradeoff is that you own uptime and result quality depends on the instance's configured engines.
claude mcp add searxng -e SEARXNG_URL=https://your-instance.example -- npx -y mcp-searxngDirectory page: SearXNG MCP.
Comparison Table
All GitHub star counts and pricing figures below were checked on July 3, 2026 against each repo and official pricing page.
| Rank | Server | GitHub stars | Free tier | Paid pricing (entry) | Best for |
|---|---|---|---|---|---|
| 1 | Tavily | 2,172 | 1,000 credits/mo | $0.008 per credit | General agent web search |
| 2 | Exa | 4,658 | Up to 20,000 requests/mo | $7 per 1,000 searches | Semantic and deep research |
| 3 | Brave Search | 1,260 | $5 credits/mo (~1,000 requests) | $5 per 1,000 requests | Independent index, low cost |
| 4 | Firecrawl | 6,796 | 1,000 credits/mo | $16/mo Hobby (yearly) | Search plus full-page scraping |
| 5 | Perplexity | 2,360 | None (pay per use) | $1/M tokens + $5 per 1,000 requests | Cited, synthesized answers |
| 6 | SerpApi | 151 | 250 searches/mo | $25/mo for 1,000 searches | Structured Google SERP data |
| 7 | Kagi | 432 | None | Usage-based, unpublished | Ad-free, quality-weighted results |
| 8 | SearXNG | 994 | Fully free | Self-hosting costs only | Zero-budget, private search |
Which Server for Which Agent Use Case
The right web search MCP server depends on what your agent does with the results, not on stars alone. This is the mapping we use when configuring agents:
| Agent use case | Pick | Why |
|---|---|---|
| Daily-driver assistant (OpenClaw, general Claude Code) | Tavily | LLM-ready results, generous free tier, one server covers search plus extract |
| Coding agent that researches libraries and APIs | Exa | Code-context search returns real usage examples, semantic matching finds docs keywords miss |
| Research agent producing cited reports | Perplexity | Sonar returns synthesized answers with citations instead of raw links |
| Scrape-heavy agent that reads full pages | Firecrawl | Search, scrape, crawl, and extract in one server and one credit pool |
| SEO, rank-tracking, or shopping agents | SerpApi | Structured Google SERP verticals with positions preserved |
| High-volume agent on a small budget | Brave Search | $5 per 1,000 requests with a recurring free credit, independent index |
| Privacy-first or homelab agent | SearXNG | No API key, no per-query billing, runs on your own hardware |
| Reading-quality-sensitive personal agent | Kagi | Ad-free, spam-downweighted results, if the usage-based cost fits |
Limitations and Tradeoffs
No web search MCP server is free at agent scale except SearXNG, and even that costs you hosting and maintenance. Three honest caveats before you pick one. First, this ranking is built from repo signals and official documentation, not a controlled head-to-head benchmark; your result quality will vary by query domain. Second, pricing drifts: Brave restructured its plans in 2026 and Exa's free allocation has changed before, so treat the numbers above as accurate for July 2026 and re-check the linked pricing pages. Third, every hosted option means your agent's queries leave your machine under that vendor's data policy, and any MCP server you add expands your attack surface, so read our guide to securing MCP server connections. If you only need occasional lookups, a plain fetch tool may be enough; see MCP servers vs API integrations for when each approach wins.
Related Guides
- OpenClaw MCP Servers Explained
- Claude Code MCP: How to Add and Manage MCP Servers
- Best MCP Servers in 2026: The Complete Ranked List
- Brave Search API for OpenClaw: Setup Guide
Go deeper
The operator playbooks
Production-ready PDF guides for OpenClaw and Hermes Agent — $19.99 each.
Skills for this topic
Browse all skills →Frequently Asked Questions
What is the best web search MCP server?
Tavily is the best web search MCP server for most AI agents in 2026. It is purpose-built for LLMs, officially maintained with 2,172 GitHub stars, and includes 1,000 free API credits per month. Exa is the better pick for semantic deep research, and Firecrawl if you also need scraping.
Which web search MCP server is free?
SearXNG is the only fully free option with no API key: the mcp-searxng server connects to any self-hosted SearXNG instance at zero per-query cost. Among hosted APIs, Tavily (1,000 credits per month), Firecrawl (1,000 credits per month), Brave Search ($5 in monthly credits), and SerpApi (250 searches per month) all have recurring free tiers as of July 2026.
What is the difference between Tavily and Exa?
Tavily is optimized for general agent web search, returning cleaned, relevance-scored snippets ready for a context window. Exa is a neural search engine that matches by meaning, with deep research and code-context tools. Use Tavily as a default search layer and Exa when queries are conceptual or research-heavy.
How do I add a web search MCP server to Claude Code?
Run claude mcp add <name> -e API_KEY=your-key -- npx -y <package> in your terminal, for example claude mcp add tavily -e TAVILY_API_KEY=your-key -- npx -y tavily-mcp . Claude Code then exposes the server's search tools in every session. The same servers work in OpenClaw, Cursor, and any MCP-compatible client.
Do AI agents need a search MCP server at all?
Only if the model or platform lacks built-in browsing, or you need a specific index. A search MCP server gives you provider choice, structured results, and predictable per-query pricing, which matters once an autonomous agent runs hundreds of searches a day. For occasional lookups, a built-in web tool or a simple fetch integration is often enough.