Best Web Search MCP Servers (2026)
Six MCP servers, one job: let an AI agent search the web without you copy-pasting from a browser tab. They cover overlapping but different shapes of the search problem — DuckDuckGo and SearXNG are the free defaults, Tavily and Exa are agent-tuned paid options, Jina is a fetch-and-extract layer, and Brave fronts an independent index. We pulled every fact from official docs, repos, and pricing pages.
On this page · 13 sections▾
TL;DR + decision tree
- Zero-budget prototype? Install DuckDuckGo MCP. No API key, no rate limit at low volumes, the most-trafficked search server on this directory at 6,333 monthly GSC impressions. It just works.
- Production agent that needs synthesis? Pay for Tavily. Returns synthesized answers plus extraction-quality content per result in one call. Built specifically for AI agents.
- Need clean Markdown of fetched pages? Pair Jina AI MCP with any search server. Reader mode at
r.jina.aiturns any URL into clean Markdown. Free + unauthenticated. - Want a non-Google index? Brave Search MCP wraps Brave’s independent crawl. Free tier covers 2K queries/month; paid plans scale predictably.
- Research-style semantic queries? Exa embeds your query and returns by similarity, not keyword match. Best when the right answer doesn’t share keywords with the question.
- Self-host required? SearXNG MCP aggregates Google, Bing, DDG, Wikipedia, and dozens more into one endpoint. Run on a $5 VPS for full privacy.
These six are not substitutes — they trade off freshness, synthesis, semantic similarity, page extraction, and self-host posture against each other. Most production agents wire two or three: a free fallback (DuckDuckGo or SearXNG), a paid synthesis layer (Tavily), and Jina for reader mode when the agent needs full page content rather than snippets.
What web search MCP servers actually do
A web search MCP server is a thin tool layer over a search provider. The agent doesn’t learn how to crawl the web — that happens at the provider (DuckDuckGo’s index, Brave’s crawl, Tavily’s pipeline, etc.). The MCP server gives the model a structured way to query that index and, in some cases, fetch the underlying pages.
The mental model is four query patterns:
- Keyword search. “Find articles about X.” Returns title + URL + snippet rows. DuckDuckGo, Brave, and SearXNG live here. Cheapest tier of any agent workflow.
- Synthesized answers. “What’s the consensus on X?” Returns an LLM-written summary plus the sources behind it. Tavily owns this category. Saves the agent from doing its own summarization.
- Semantic search. “Find content conceptually related to X.” Returns results ranked by embedding similarity, not keyword match. Exa owns this. Best for research-style queries where the right answer doesn’t share words with the question.
- Page fetch + extract. “Read this URL as clean Markdown.” Jina’s reader mode owns this. Pairs with any search server — search returns URLs, Jina turns them into structured content.
A well-behaved web search MCP keeps tool descriptions short, exposes pagination so the model doesn’t blow context on a 10k-result response, and surfaces rate-limit metadata so agents can back off cleanly. If you’re new to the protocol underneath, our What is MCP primer covers the JSON-RPC wire format these servers run on.
Side-by-side matrix
Every cell is sourced from official docs or vendor pricing pages. Live counts checked 2026-05-08 from indexed entries on this directory.
| Dimension | DuckDuckGo | Tavily | Jina | Brave | Exa | SearXNG |
|---|---|---|---|---|---|---|
| License (MCP) | MIT (community) | MIT | MIT | MIT | MIT | MIT (SearXNG: AGPL) |
| API key required | No | Yes | Reader: no / Search: yes | Yes | Yes | No (public) / yes if self-hosted |
| Free tier | Unlimited (no key) | 1,000 calls/mo | Reader free; search free tier | 2,000 queries/mo | Small free tier | Unlimited (self-host) |
| Paid pricing | N/A — free | Usage-based, scales linearly | Usage-based | Usage-based per 1K queries | Usage-based per query | N/A — your VPS bill |
| Search type | Keyword | Synthesized + keyword | Reader (page fetch) | Keyword | Semantic (neural) | Keyword (meta-aggregator) |
| Self-host option | No (DDG itself) | No | No | No | No | Yes — fully |
| MCP.Directory page | /servers/duckduckgo | tavily.com (no MCP listing) | /servers/jina-ai | /servers/brave-search | exa.ai (no MCP listing) | /servers/searxng-public |
| Best for | Zero-budget prototypes | Production agent synthesis | Reading full page content | Independent crawl, non-Google | Research, conceptual queries | Privacy + multi-engine aggregation |
Three takeaways. First, DuckDuckGo and SearXNG are the only free-forever options — every other entry meters past a small free tier. Second, Tavily and Exa are the only servers built specifically for AI agents — the others wrap a general-purpose search API. Third, Jina is not a search engine — it’s a fetch-and-extract layer that pairs with any of the others. Don’t pick Jina vs DuckDuckGo; you usually want both.
DuckDuckGo MCP — install + recipe
What it does best
DuckDuckGo MCP is the default “I just want web search to work” option. Free, no API key, the highest-trafficked search server on this directory at 6,333 monthly GSC impressions. Returns clean DDG results — no ads, no tracking. The community-maintained server wraps DuckDuckGo’s Instant Answer API plus the HTML SERP scrape. Latency is moderate (300-800ms per query) but fully free with no rate-limit visible at low query volumes. The 30-second install + zero credentials makes it the right starting point for any new agent project.
Pick this if you...
- Have a zero-budget agent that just needs basic web search
- Are building a quick prototype and want sub-minute setup
- Want privacy-first search — DuckDuckGo doesn’t track queries
- Only need keyword search, not semantic similarity or synthesis
Recipe: classify the top 5 results for a query
In any MCP client (Cursor, Claude Code, VS Code, Claude Desktop) with DuckDuckGo MCP installed:
Search DuckDuckGo for "context7 alternatives 2026" and return
the top 5 results with title + URL + snippet. For each result,
indicate whether it's a docs site, a blog post, or a marketplace
listing.The agent calls the duckduckgo MCP’s search tool with the query, gets back ~10 results, classifies each by URL pattern (docs.* / blog/* / marketplace/*), and returns the top 5. Total cost: $0 plus whatever the model itself charges. Latency: typically under 1 second.
Skip it if...
You need agent-tuned synthesized answers (Tavily wins here); you need clean Markdown of fetched pages (Jina wins); you need semantic similarity search (Exa wins). DuckDuckGo gives you raw keyword-matched SERP results — anything beyond that is your agent’s job.
Tavily MCP — install + recipe
Tavily — not currently listed on mcp.directory; visit the Tavily site for install + API keys.
What it does best
Tavily is the agent-tuned paid option. Their API returns synthesized answers (a short LLM-written summary) plus the raw results in one call — saves the agent from doing its own summarization. Free tier covers 1,000 calls/month; paid plans scale linearly. The Tavily team built it specifically for AI agents, which shows in the API design (returns include a content field with extraction-quality content per result, not just snippets). Latency is competitive with DuckDuckGo despite the synthesis pass.
Pick this if you...
- Are building a production agent that needs reliable synthesis
- Are willing to pay $0.005-0.01 per query for higher-quality output
- Care about extraction-quality content (not just snippets) per result
- Want a single-call answer + sources pattern for citation workflows
Recipe: research a topic with answer + sources
In any MCP client with the Tavily MCP installed and your Tavily API key set:
Use Tavily to research "OAuth 2.1 PKCE flow for MCP servers"
— return Tavily's synthesized answer plus the top 5 sources
with extraction-quality content for each.The agent calls Tavily’s search endpoint with include_answer=true&max_results=5, receives a synthesized answer plus 5 results with full content fields, and writes a 200-word summary citing each. The synthesis pass replaces a separate model round-trip the agent would otherwise need to do over raw snippets, which is where Tavily’s money saves you money.
Skip it if...
You have zero budget — Tavily’s free tier is small. You need keyword-first search and don’t want synthesis getting in the way (Tavily over-indexes on synthesis). You’re building anything where DuckDuckGo’s free tier is enough — no need to pay for capability you don’t use.
Jina AI MCP — install + recipe
What it does best
Jina AI’s reader-mode MCP turns any URL into clean Markdown. It’s not a search engine itself — it’s a fetch-and-extract layer that pairs with web search. Reader mode at r.jina.ai is free and unauthenticated; the search API at s.jina.ai is the search side. Best when the agent needs the actual content of pages, not just titles + snippets. Caches aggressively, so repeated fetches of the same URL are cheap. Renders JavaScript before extracting, which matters for SPAs and dynamic dashboards.
Pick this if you...
- Need agents to read full content of search results, not just snippets
- Want free reader mode for any URL with no API key required
- Are building a research agent that summarizes long articles
- Care about JavaScript-rendered pages — Jina renders before extracting
Recipe: read three URLs as clean Markdown
In any MCP client with Jina MCP installed:
Use Jina to fetch and read these 3 URLs as clean Markdown:
blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps
latent.space/p/why-mcp-won
simonwillison.net/2025/Oct/16/claude-skills/
Summarize each in 100 words.Agent calls Jina’s reader endpoint for each URL, receives clean Markdown, summarizes each. The Markdown is much cheaper to process than raw HTML — typically 70-90% fewer tokens than the original page, with all navigation, ads, and boilerplate stripped. The same prompt shape works for any URL the agent encounters; pair Jina with DuckDuckGo or Brave to make a search-then-read agent.
Skip it if...
You only need snippets — DuckDuckGo or Brave is faster and cheaper. You’re scraping JavaScript-heavy SPAs that Jina’s renderer struggles with — try a Playwright-based MCP instead. You need to scale beyond Jina’s free tier for reader mode (paid plans exist for the search API; check current pricing).
Brave Search MCP — install + recipe
What it does best
Brave Search MCP wraps Brave’s independent index — a real alternative to Google. Modest free tier (2,000 queries/month at the time of writing), then paid. The index is built from Brave’s own web crawl, not licensed from Google or Bing — different result distribution, sometimes better for non-English / non-US queries. Privacy posture similar to DuckDuckGo. The MCP returns standard SERP-shaped results with predictable per-1K-query pricing past the free tier, which makes it a sensible production option.
Pick this if you...
- Want a non-Google index — different results, less SEO-spam-dominated
- Need a paid tier with predictable pricing for production agents
- Care about privacy plus an independent crawl (no upstream licensing)
- Are building agents that target non-US users — Brave’s crawl is more globally distributed than DDG’s
Recipe: find docs over blog posts for a query
In any MCP client with Brave Search MCP installed and your BRAVE_API_KEY set:
Use Brave Search to find documentation pages for "MCP OAuth
2.1 implementation" — return the top 5 results, prioritize
official spec pages over blog posts. Compare against the same
query on DuckDuckGo.Agent calls Brave’s search endpoint, filters for results matching *.modelcontextprotocol.io or spec.*, returns top 5. Cross-checks against DuckDuckGo for distribution comparison — gives you a real read on whether your query has consistent ranking across independent indexes. Useful for both production routing logic and as a sanity check during agent development.
Skip it if...
Your budget is zero and you’ll exceed Brave’s free tier — DuckDuckGo or self-hosted SearXNG cover that floor. You need synthesized answers (Tavily wins). You’re already on a paid Bing or Google API and don’t need another index for cross-checking.
Exa MCP — install + recipe
Exa — not currently listed on mcp.directory; visit the Exa site for install + API keys.
What it does best
Exa (formerly Metaphor) is the semantic search option. Instead of keyword-matching, Exa embeds your query and returns results by semantic similarity. Best for research-style queries where the right answer doesn’t share keywords with the question. Paid API; free tier is small. Returns include a similarity score that’s actually useful — agents can filter by score threshold to reject conceptually-distant results before passing them back to the model.
Pick this if you...
- Need semantic search — “how do agents handle context bloat” should return results about progressive disclosure even without keyword matches
- Are building research agents that explore an idea space
- Want similarity scores for ranking + filtering
- Are willing to pay per query for higher-quality research-mode results
Recipe: semantic search with score-based filtering
In any MCP client with Exa MCP installed and your EXA_API_KEY set:
Use Exa semantic search to find articles about "agents handling
tool surface area without overwhelming the context window."
Return the top 10 results sorted by similarity score; include
the score for each.Agent calls Exa’s /search endpoint with type=neural, gets back results ranked by embedding similarity to the query, returns top 10. The scores let the agent reject irrelevant results (similarity < 0.7), which keyword search can’t do without a separate ranking pass. This is the Exa value-add: query semantics in, semantically-filtered results out, no hand-tuned filter logic on the agent side.
Skip it if...
Keyword search is fine — DuckDuckGo and Brave are cheaper and faster. Zero budget — Exa’s free tier is small. You’re building a fact-lookup agent — semantic search returns conceptually-related results that may not contain the specific fact you need.
SearXNG MCP — install + recipe
What it does best
SearXNG is the self-hosted meta-search option. It aggregates Google, Bing, DuckDuckGo, Wikipedia, and dozens of other engines into a single response. Open source, runs on a $5 VPS or your own machine. The mcp.directory server (1,468 GSC impressions) wraps a public SearXNG instance; you can also point it at your own. Best privacy posture — no provider sees your query if you self-host. Best for catching results no single engine has, since the meta-aggregation pulls from many providers in parallel.
Pick this if you...
- Need to aggregate results across many engines — catches results no single engine has
- Want full self-hosting for privacy compliance
- Run agents in regulated environments (healthcare, finance, legal) where query data must not leave your network
- Want to bypass single-provider rate limits by spreading query load
Recipe: aggregate distinct results across providers
In any MCP client with SearXNG MCP installed and pointed at a SearXNG instance (public or your own):
Use SearXNG to aggregate web search across Google, Bing,
DuckDuckGo, and Wikipedia for "MCP context bloat 2026."
Return distinct results from each source, indicate which
provider each came from.Agent calls SearXNG with engines=google,bing,duckduckgo,wikipedia, gets a deduplicated result set with provider metadata per result, returns top 5 distinct results showing which provider hit them. The provider tags let you reason about index disagreement — when a result only appears in one engine, you know to weight it differently.
Skip it if...
Managing your own SearXNG instance is operational overhead you don’t want — running one is a cron job and a health check, not nothing. Public SearXNG instances rate-limit aggressively (set up your own if you hit the wall). Single-provider results from DuckDuckGo or Brave are good enough — don’t pay the aggregation tax if you don’t need it.
Common pitfalls
DuckDuckGo throttling on hot loops
DuckDuckGo’s rate limit is invisible until you hit it. A model in a tight loop firing 100+ queries per minute will get HTML 403s back. Cap the agent’s tool-call budget per turn, batch related queries into one search, and prefer one well-scoped query over many narrow ones. The community DuckDuckGo MCP surfaces 403s cleanly — handle them in your client.
Tavily synthesis vs raw sources confusion
Tavily returns both an answer field and results[]. Agents that cite the synthesized answer as a source instead of citing the underlying URLs produce hallucinated-looking citations. Prompt explicitly: “cite results[].url, not the synthesized answer.” Or set include_answer=false when you only want raw results.
Jina reader on SPA-heavy pages
Jina renders JavaScript before extracting, but heavy SPAs (React-heavy dashboards, infinite-scroll feeds) sometimes return partial Markdown. If you see suspiciously short content from Jina on a page that should be long, fall back to a Playwright-based scraper or fetch the page’s underlying API directly.
Brave free-tier surprise bills
The Brave free tier is 2K queries/month, but production agents looping on a busy site can blow through that in a day. Track query count per agent run and alert before you hit the cap. Brave’s API returns rate-limit headers — surface them through the MCP and let your client gate.
Exa keyword fallback expectation
Exa is semantic-first. If you ask for “the page titled X” expecting a keyword match, you may get pages that are conceptually similar to X but don’t contain the exact title. Set type=keyword for exact-match queries, or use Exa’s find-similar feature for “more like this URL.”
Six at once = description bloat
Each of these servers exposes 3-10 tools. Stacking all six in one client adds 30+ tool descriptions to every prompt. Pick 2-3 that match the question shape — usually one keyword (DuckDuckGo or Brave), one synthesis (Tavily), one reader (Jina). For more on this trade-off, see our MCP context bloat post.
Community signal
The web search MCP space is more mature than the observability MCP space — DuckDuckGo and Brave wrappers shipped early in 2025, and Tavily and Exa already had agent-friendly APIs from before MCP existed. That means more production-track-record per server, more StackOverflow-grade Q&A on rate limits, and clearer pricing signals.
What’s consistent across HN threads, GitHub discussions, and vendor blogs in the lead-up to 2026: agents that wire one search MCP underperform agents that wire 2-3 with explicit routing. The pattern that wins is: DuckDuckGo or SearXNG as the always-available fallback, Tavily for the synthesized-answer + extraction-content pattern in production, Jina as the reader-mode pair when the agent needs full page content. Exa shows up in research-agent and discovery-agent stacks where keyword search misses the concept-shaped queries; it’s less common in fact-lookup agents. Brave is gaining traction as the non-Google paid index of choice — predictable pricing plus an independent crawl plays well with privacy-conscious buyers. The DuckDuckGo MCP’s 6,333 monthly impressions on this directory is the single biggest signal in the cluster.
Frequently asked questions
Best free web search MCP server in 2026?
DuckDuckGo MCP is the unequivocal winner — no API key, unlimited queries, zero cost, the most-trafficked search server on this directory. SearXNG is a strong second if you self-host. Brave Search MCP has a free tier (2K queries/month) that beats DuckDuckGo on result quality but starts charging at scale.
DuckDuckGo MCP vs Tavily — which is better?
Different jobs. DuckDuckGo gives raw web results — perfect for prototypes and zero-budget agents. Tavily gives synthesized answers + extraction-quality content per result, optimized for AI agents in production. If you'd rather pay $5-50/month and ship a higher-quality agent, Tavily. If budget is zero, DuckDuckGo.
Is there a Brave Search MCP server?
Yes — at /servers/brave-search on this directory. Wraps Brave's independent search index. Free tier covers 2,000 queries/month at the time of writing; paid plans scale predictably. Different result distribution than Google or DuckDuckGo because Brave's index is built from their own crawl.
Can I self-host web search MCP?
SearXNG is the self-host answer — open source, runs on any VPS, aggregates dozens of search engines without provider lock-in. The mcp.directory searxng-public server wraps a public instance; for production you'd point it at your own SearXNG deployment. DuckDuckGo, Brave, Tavily, and Exa are all hosted-only.
Tavily vs Exa — which is more agent-friendly?
Tavily for fact-lookup tasks where the agent needs a quick answer + sources. Exa for research tasks where the agent is exploring an idea space and keyword search misses semantically-related content. Many production agents use both: Tavily for synthesized answer-mode, Exa for 'find me everything related to this concept.'
How do I add a web search MCP to Claude Code?
Open the install card on any of the server pages (e.g., /servers/duckduckgo) — it generates a copy-pasteable JSON config block for Claude Code, Cursor, VS Code, Codex, Gemini, Windsurf, and ChatGPT. Pick your client tab, copy the command, paste into your client's MCP config, restart the client. The server appears as a tool in your next prompt.
Does DuckDuckGo MCP need an API key?
No. The community-maintained DuckDuckGo MCP wraps DDG's HTML interface and Instant Answer API, both of which are unauthenticated. Set up takes ~30 seconds — copy the install command, restart your MCP client, done. The trade-off is rate-limit visibility — if you query DDG hundreds of times per minute you may get throttled, but for normal agent use the limit is invisible.
Best web search MCP for production agents?
Tavily for the synthesized-answer + extraction-content pattern — it's purpose-built for agents and the API surface reflects that. Exa as the secondary for semantic queries. DuckDuckGo or SearXNG as the always-available fallback when paid quotas are exhausted. Most production agents wire 2-3 search MCPs and route based on query intent.
Sources
DuckDuckGo
- /servers/duckduckgo — install card + tools list
- duckduckgo.com/api — Instant Answer API docs
Tavily
- tavily.com — official site, API + pricing
- docs.tavily.com — search endpoint reference
Jina AI
- /servers/jina-ai — install card
- jina.ai/reader — reader-mode docs (r.jina.ai)
- s.jina.ai — search endpoint
Brave Search
- /servers/brave-search — install card
- brave.com/search/api — API docs + pricing tiers
Exa
- exa.ai — official site (formerly Metaphor)
- docs.exa.ai — neural search endpoint reference
SearXNG
- /servers/searxng-public — install card
- docs.searxng.org — self-host docs, supported engines
- github.com/searxng/searxng — AGPL-licensed source
Related comparisons
- /blog/chrome-devtools-mcp-vs-playwright-mcp-2026 — browser automation, related to scraping
- /blog/firecrawl-vs-anycrawl-vs-crawleo-vs-playwright-2026 (placeholder — post in progress) — scraper deep dive (companion piece)
- /blog/best-language-skills-for-claude-code-2026 — sibling D-1 comparison
- /blog/datadog-vs-grafana-vs-sentry-mcp-2026 — observability deep dive
Internal links
- /servers/duckduckgo
- /servers/brave-search
- /servers/jina-ai
- /servers/searxng-public
- /blog/what-is-mcp — protocol primer
- /best-mcp-servers — curated roundup
- /servers — browse all 3,000+
The pragmatic stack
Most production agents in 2026 wire three of these: DuckDuckGo or SearXNG as the always-free fallback, Tavily for synthesized-answer queries, and Jina for reader-mode page extraction. Add Brave or Exa when you hit a category-specific need — non-Google index or semantic similarity respectively. Watch the tool-description budget — six search MCPs in one client is too many.
Keep reading
Comparison
Chrome DevTools MCP vs Playwright MCP
ReadComparison · placeholder
Firecrawl vs AnyCrawl vs Crawleo vs Playwright
Post in progressComparison