Ranked roundup · Web search
Best MCP Server for Web Search
Picking an MCP server for web search is less about finding one that works and more about choosing between five official ones that do. Every major search-API vendor now ships a first-party server — Tavily, Exa, Perplexity, Brave, Firecrawl, and Kagi all maintain official MCP repos, and we checked each against the GitHub API on September 1, 2026. What's left for the community is the free-and-keyless niche, which a DuckDuckGo scraper with about 1,400 stars fills adequately, caveats included. The rankings below use verifiable criteria only: official vendor support, commit recency, adoption, auth model, and what the tools cover per each vendor's own docs. Nobody paid for placement, and we ran no benchmarks — where we couldn't verify a claim, we say so. Two things changed this year that stale roundups still miss: Brave dropped its no-card free tier in February 2026, and the original brave-search reference server sits in an archived repo. Both are covered below.
Facts checked September 1, 2026 · ranked on verifiable criteria · placement is never for sale
Founder, Automation Squad ·
The ranking
- Official · Tavily
Agent builders who want search plus content retrieval behind one key, with a real monthly free tier (1,000 credits, no card)
Four tools: real-time web search, page extract, site map, and crawl — the widest search-first coverage of any official server
First-party hosted remote endpoint (mcp.tavily.com/mcp) or local via npx · API key or OAuth; hosted endpoint accepts the key via header or URL parameter — use the header or OAuth, not the URL · MIT · 2,365★
- Free tier is 1,000 credits/month and advanced searches burn 2 credits each — heavy agents will hit the ceiling fast; paid plans start at $30/month
- The README shows passing your API key as a URL parameter, which leaks keys into logs — use the Authorization header or OAuth instead
- Official · Exa
Trying MCP web search in under a minute — paste the hosted URL into your client and it works before you have an account
Web search with clean content extraction, full-page fetch as markdown; optional deep-research agent and advanced filtered search behind URL flags
First-party hosted remote endpoint (mcp.exa.ai/mcp); local install also available · Anonymous (rate-limited), OAuth, or API key — the only official server usable with zero signup · MIT · 4,951★
- Anonymous access is rate-limited and Exa doesn't publish the limits — fine for evaluation, not for production
- The deep-research and advanced-search tools are off by default and need URL parameters to enable
- Paid usage is metered pay-as-you-go; we could not verify a documented recurring free credit allowance beyond the anonymous tier
- Official · Firecrawl
Workflows where finding pages is step one and getting their full content is the actual job
Search is one tool in a scraping suite: scrape, crawl, map, and structured extract — the strongest fetch-and-crawl coverage on this page
Local via npx/Docker or Firecrawl's hosted option; self-hostable since the core engine is open source · API key from the Firecrawl dashboard · MIT · 7,362★
- Search is secondary to scraping here — if you only need SERP-style results, Tavily or Exa are more direct
- Credits are per-page, don't roll over, and the free plan is 1,000 credits/month (no card) per the current pricing page — third-party posts still cite an older 500-credit cap
- Highest-volume repo on this list, but that reflects the whole Firecrawl toolset, not web search specifically
- Official · Perplexity
Agents that want synthesized, cited answers and multi-step research rather than raw result lists
Four tools spanning ranked web search, conversational ask, deep research, and reasoning — the only server here with an answer-engine layer built in
First-party hosted remote endpoint (api.perplexity.ai/mcp) or local install · API key from console.perplexity.ai; prepaid credit balance with a payment method on file · MIT · 2,495★
- No free tier: the API requires prepaid credits and a payment method before the first call
- Raw search runs $5 per 1,000 requests; the ask/research/reason tools bill LLM token costs on top
- Often mislabeled a community server in older roundups — it's official, but the repo name (modelcontextprotocol) makes it easy to miss
- Official · Brave
Teams that want results from an independent index (no Google/Bing backend) with news and local verticals in one server
Eight tools: web, news, image, video, local/place search, LLM-ready context, and summarizer — the broadest vertical coverage here, backed by Brave's independent index
Self-host only — npx or Docker, STDIO or HTTP transport; no first-party hosted MCP endpoint · API key from Brave's developer dashboard; a card on file is now required for every plan · MIT · 1,411★
- The no-card free tier died in February 2026 — every plan now requires a card and grants $5/month in credits (roughly 1,000 searches) before billing at $5 per 1,000 requests
- No hosted MCP endpoint, so you're running it yourself
- Older tutorials point at the archived reference implementation, not this repo
- Community
Hobby projects and local agents where free-and-keyless matters more than uptime guarantees
Web search through DuckDuckGo's public results plus content fetching and parsing, with built-in rate limiting
Local only (stdio), install via uv/pip · None — no API key, no account · MIT · 1,446★
- Community project with no affiliation to DuckDuckGo — no API contract, no SLA, no support
- It parses public result pages, so it can be rate-limited or silently broken by upstream changes at any time
- Slower maintenance cadence than the official servers (last push July 28, 2026, vs. late August for every vendor repo)
Web search MCP servers compared
| Server | Maintained by | Runs | Auth | License |
|---|---|---|---|---|
| Tavily MCP Server | Tavily | First-party hosted remote endpoint (mcp.tavily.com/mcp) or local via npx | API key or OAuth; hosted endpoint accepts the key via header or URL parameter — use the header or OAuth, not the URL | MIT |
| Exa MCP Server | Exa | First-party hosted remote endpoint (mcp.exa.ai/mcp); local install also available | Anonymous (rate-limited), OAuth, or API key — the only official server usable with zero signup | MIT |
| Firecrawl MCP Server | Firecrawl | Local via npx/Docker or Firecrawl's hosted option; self-hostable since the core engine is open source | API key from the Firecrawl dashboard | MIT |
| Perplexity MCP Server | Perplexity | First-party hosted remote endpoint (api.perplexity.ai/mcp) or local install | API key from console.perplexity.ai; prepaid credit balance with a payment method on file | MIT |
| Brave Search MCP Server | Brave | Self-host only — npx or Docker, STDIO or HTTP transport; no first-party hosted MCP endpoint | API key from Brave's developer dashboard; a card on file is now required for every plan | MIT |
| DuckDuckGo MCP Server (community) | Community | Local only (stdio), install via uv/pip | None — no API key, no account | MIT |
How to choose
Pick Tavily if you want agent-tuned search with a usable free tier (1,000 credits a month, no card) plus extract, map, and crawl in one server. Pick Exa if you want the lowest friction anywhere — its hosted endpoint works anonymously, rate-limited, before you make an account. Pick Firecrawl if scraping and crawling matter more than the search itself. Pick Perplexity if you want answer-style search and deep research and accept prepaid billing with no free tier. Pick Brave for an independent index with news, image, video, and local verticals — accepting a card on file. Pick the DuckDuckGo community server only when free-and-keyless outranks reliability.
How this ranking works
Every server here is real and was checked against its repository or vendor documentation on the date above — stars and licenses come from the GitHub API, not from memory. Order reflects verifiable criteria only: official support, maintenance recency, adoption, auth model, and documented capability coverage. We do not run paid placements, and we list every caveat we find, including the unflattering ones.
Common questions
- Is there an official MCP server for web search?
- There are six: Tavily, Exa, Perplexity, Brave, Firecrawl, and Kagi all maintain first-party MCP servers in their own GitHub orgs. This category doesn't need community stand-ins for the major providers — the community's real contribution is the keyless DuckDuckGo option.
- What does setup and auth actually look like?
- Exa's hosted endpoint works anonymously with rate limits; Tavily's hosted endpoint takes OAuth or an API key; Perplexity's takes an API key backed by prepaid credits. Brave and DuckDuckGo are local installs via npx, Docker, or pip. Every option except DuckDuckGo eventually needs an API key — treat keys like credentials and pass them in headers or via OAuth, not pasted into URLs.
- Can I run web search over MCP completely free?
- Yes, three ways: the DuckDuckGo community server is keyless but fragile, Tavily gives 1,000 credits a month with no card, and Firecrawl gives 1,000 credits a month with no card. Exa's anonymous hosted endpoint also works free within unpublished rate limits. Brave and Perplexity both require a payment method before meaningful use.
- Why do old tutorials point at a different Brave server?
- The original brave-search reference server now lives in modelcontextprotocol/servers-archived, unmaintained since May 2025. Brave maintains its own official replacement at brave/brave-search-mcp-server — and note the API behind it dropped its no-card free tier in February 2026.
MCP servers we build
Automation Squad publishes its own open-source MCP servers and agent skills, documented with the same rules this ranking uses.