Exa MCP vs Tavily MCP vs Apify MCP: Best Search Skill

Three MCP servers keep showing up in the same breath whenever someone asks how to give Claude Code web access. They solve three different problems, and installing the wrong one produces the same symptom every time: your agent gets results, they look fine, and the answer is subtly wrong.

Exa searches by meaning. Tavily searches for things you're going to cite. Apify doesn't search at all, it extracts. Here's how each behaves once it's wired into Claude Code, and which one belongs in your config.

Head to head

Exa MCP — when you don't know the keyword

Install:

claude mcp add exa -- npx -y exa-mcp-server

Published by exa-labs from github.com/exa-labs/exa-mcp-server. MIT licensed, verified tier, community-reviewed.

Exa's search index is embedding-based rather than keyword-based, which means it answers queries phrased as descriptions instead of search terms. "Blog posts by engineers who left large companies to build database tooling" returns useful results. Google returns a listicle about leaving your job.

That's the whole pitch and it's a real one. When your agent is doing discovery rather than lookup, keyword search wastes turns. The agent issues a query, gets generic SEO pages back, reformulates, tries again. Exa short-circuits that loop because the query never had to be reverse-engineered into keywords.

Where it disappoints: recency. Semantic indexes lag. If you ask about something that happened this week, Exa often returns thematically-correct results from eighteen months ago, presented with the same confidence as fresh ones. For anything time-sensitive, this is the failure mode that bites.

The free tier gives 1,000 searches a month, then $100/mo Growth. For an agent that searches on most turns, 1,000 goes faster than you'd expect.

Tavily MCP — built for the citation

Install:

claude mcp add tavily -- npx -y tavily-mcp

Published by tavily-ai from github.com/tavily-ai/tavily-mcp. MIT, verified tier, community-reviewed.

Tavily describes itself as a production-ready MCP server with real-time search, extract, map, and crawl. The distinction that matters is extract: Tavily doesn't hand your agent a list of links and let it fetch them one by one. It returns the page content alongside the result, already cleaned up.

That single design decision changes agent behavior more than any benchmark. Without extraction, a search step becomes search plus three or four fetches plus HTML parsing, and each fetch is a chance to hit a paywall, a bot check, or 40KB of navigation markup that eats context. Tavily collapses that into one call with the text already usable.

It's also the strongest of the three on recency. Real-time search is the marketing phrase and it holds up. If your agent answers questions about current events, pricing, or anything that changed this quarter, Tavily is the one to trust.

The cost jump is the downside. Free covers 1,000 searches a month, and the next tier is $200/mo Scale, double Exa's Growth plan. There's no gentle middle step.

Apify MCP — not a search engine

Install:

claude mcp add apify -- npx -y @apify/actors-mcp-server

Published by Apify from github.com/apify/apify-mcp-server. MIT, verified tier, community-reviewed.

Apify exposes its Actor library to your agent. There are over 20,000 of them, prewritten scrapers for specific sites, running on managed proxy infrastructure. Instagram profiles, Amazon listings, Google Maps results, LinkedIn pages, TripAdvisor reviews. The agent picks the Actor and gets structured JSON back.

Calling this a search tool undersells what it does and oversells what it's for. If you ask Apify a question, you'll be disappointed. If you tell it to pull every review for a set of 200 businesses with rating, date, and text as separate fields, it does that and neither Exa nor Tavily can.

The proxy layer is why people pay. Scraping Instagram or Amazon from your own IP gets you blocked in minutes. Apify handles rotation, and the $5 free monthly credit is enough to test whether an Actor works for your case before committing.

Pricing is credit-based, which makes it the hardest of the three to forecast. A heavy Actor on a large target burns credits fast. Starter is $29/mo, Scale $199/mo.

Which wins on four real tasks

"Find me prior art for this technical approach." Exa. This is exactly the query type keyword search handles badly, and freshness doesn't matter for prior art.

"What's the current pricing for these five SaaS products?" Tavily. Recency plus extraction in one call. Exa will confidently return a pricing page from 2024.

"Build a dataset of every listing in this category with price and seller." Apify. The other two return prose about listings. Apify returns the listings.

"Research this company before my meeting." Tavily first for the recent news, then Exa for the deeper background pieces that don't rank. Running both here is reasonable and the combined cost is still less than one hour of doing it manually.

Cost, honestly

All three publish a 1,000-request-equivalent free tier and all three are consumed faster than people plan for. An agent doing autonomous research issues three to eight searches per task. At five per task, 1,000 searches is 200 tasks, which is roughly a month of moderate use by one developer.

The trap is running search on autopilot inside a loop. An agent that re-searches on every iteration of a five-step plan will exhaust a monthly tier in a weekend. Set a budget cap on the API side, not just in your head.

If you want one paid subscription rather than three, the ranking depends on your work. Research-heavy and cost-sensitive: Exa at $100. Anything user-facing where a wrong date is unacceptable: Tavily at $200. Data collection: Apify at $29, which is the best value on this list by a wide margin, for a narrower job.

Running more than one

MCP Omnisearch is worth knowing about here. It's a community-tier server that puts Tavily, Brave, Kagi, and Exa behind one interface. MIT, unreviewed security status.

The appeal is obvious and so is the risk. One config block instead of four, but you're now depending on an unreviewed single-maintainer project sitting between your agent and every API key you own. For a personal setup, fine. For anything touching client data, install the first-party servers directly.

Two other options if these three don't fit. Firecrawl, installed with npx -y firecrawl-mcp, handles JS-heavy pages and full-site crawls better than any of them, and we compared it against the deep research approach in Deep Research Skill vs Firecrawl. LLM Scraper is a library rather than a server, for people who want the extraction logic in their own code.

Security notes before you install

All three of these are first-party servers published by the company that runs the API, which is the single best signal available in the MCP ecosystem right now. All three carry community-reviewed status and MIT licenses.

Two things to still check. First, the npx -y install pattern pulls and executes the package at every start; pin a version in production rather than tracking latest. Second, every one of these servers holds an API key with billing attached, so treat the config file as a secret and keep it out of any repo you push.

The broader risk with search MCP servers isn't the server, it's what comes back through it. Content fetched from the web enters your agent's context as text, and text can contain instructions. Our writeup on web scraping agent skills goes into how to bound that.

The recommendation

Install Tavily if you install exactly one. Extraction plus recency covers the largest share of what people ask agents to do on the web, and the failure mode of stale-but-confident results is worse than the failure mode of paying more.

Add Exa when your work is research rather than lookup, and you'll notice the difference within a day.

Add Apify only when you have a specific structured extraction job. It's excellent at that job and useless outside it, and the credit model punishes casual use.

FAQ

Can I run all three at once in Claude Code? Yes, MCP servers coexist and Claude picks based on the tool descriptions. The practical cost is context: each server registers its tools in every prompt. Three search servers plus your other MCP tooling adds up, and on long sessions that overhead is real.

Do these work with Cursor and VS Code Copilot? All three list Claude Code, Cursor, and VS Code Copilot as compatible agents. The install command differs by client but the servers themselves are agent-agnostic.

Is Exa or Tavily better for RAG? Tavily, and it isn't close. Extraction returns clean text ready to chunk, and freshness matters more in retrieval than most people account for. Exa is better at the discovery step that happens before you build the index.

What happens when I hit the free tier limit? The API returns an error and your agent sees a failed tool call. Claude usually handles this by trying something else rather than stopping, which sounds good and isn't, since you may not notice search silently stopped working. Watch your usage dashboard.

Is Apify worth it if I already have Firecrawl? Only if your targets are sites with prewritten Actors. Firecrawl handles arbitrary pages well. Apify wins on the specific platforms where scraping is adversarial, Instagram and Amazon being the clearest examples, because you're renting their proxy infrastructure along with the scraper.