We checked Exa's MCP server against its own current README instead of trusting cached summaries. Worth doing. Two of the tool names floating around aggregator sites turned out to be stale. What's actually shipping is a search server built for AI consumption from the ground up, not a web-search API that bolted an MCP interface on later. The official server comes in two forms, a local npm package and a hosted remote endpoint. Below: what the tools do today, how the two deployment paths differ, and where Exa lands next to Perplexity's MCP server, which we've already covered.
What Is the Exa MCP Server?
Exa MCP is the official server for Exa's search API, maintained under the exa-labs GitHub org. Our directory indexes it as ai-exa-exa at 4,083 GitHub stars as of the last sync. The live count on GitHub runs a little higher, which is normal sync lag on a repo moving this fast. License is MIT.
That number sits well above the average for our search category, where 1,631 servers average 17.5 stars each. Most of that category is thin community wrappers around a single API. Exa is one of the exceptions. Perplexity's is the other, and the two are worth putting side by side, since they chase the same "give my agent live web results" job from opposite ends.
There are two deployment paths. Run the package locally over stdio with an Exa API key, or point a client at the hosted remote endpoint and authenticate with OAuth. Most MCP servers pick one. Exa ships both.
The index underneath is built differently from a general-purpose crawler, too. A traditional search engine indexes pages. Exa trains its ranking on embeddings, so a query comes back ordered by semantic similarity to the query itself, not keyword overlap. Same mechanism as Exa's standalone product. In practice that means results read as more relevant for research-style questions and worse for exact-phrase lookups, which is a tradeoff you feel the first time you need a literal string.
What Tools Does the Exa MCP Server Expose?
Four tools ship with the current release. We checked that list against the repo's own README, not a third-party cache, because one older fork still advertises a tool the canonical server dropped. Two are on by default. The other two are optional, and you have to enable them explicitly in your client config.
web_search_exa is the default general search: query in, ranked results out, scored by Exa's own relevance model rather than raw keyword matching. web_fetch_exa is on by default too. It pulls full page content for a specific URL the agent already has, which is what you want when the search result alone doesn't carry enough context to answer the question.
The optional pair handles narrower jobs. web_search_advanced_exa bolts filtering onto the default search, so an agent can pin results to a domain, a date range, or a content type and stop sifting through an unfiltered list. agent_run is a different animal. It fires off a multi-step research task inside Exa's own infrastructure and hands back a synthesized result. That's delegation, not a search call.
How Do You Install the Exa MCP Server?
The local path runs through npx, same as most MCP servers that ship as npm packages.
{
"mcpServers": {
"exa": {
"command": "npx",
"args": ["-y", "exa-mcp-server"],
"env": {
"EXA_API_KEY": "your-api-key-here"
}
}
}
}Grab the API key from your Exa dashboard first. Skip that and the server still starts, but every tool call fails. Exa's free tier includes a limited number of monthly search credits, enough to kick the tires before you decide whether to put a card down. One caveat on credits: agent_run burns through them faster than a single web_search_exa call, since each run triggers several searches internally.
The hosted path skips the local install completely. Point any MCP client that supports remote Streamable HTTP servers at mcp.exa.ai/mcp. No env var to set, because the client walks you through an OAuth flow. Both paths expose the same four tools, so the decision comes down to one question: manage a key yourself, or hand auth off to OAuth?
How Does Exa MCP Compare to Perplexity's MCP Server?
Both are official. Both are search-first. We've already covered Perplexity's MCP server in its own guide, so here's where the two actually diverge.
Perplexity's server leans on its Sonar models to hand back a synthesized, cited answer. That's closer to asking a research assistant a question than running a raw search. Exa's defaults return ranked results and page content and leave synthesis to the calling agent, unless you reach for agent_run and get Exa's own multi-step research behavior. So the real question is where synthesis should live: inside your agent, or inside the server. Exa bets on the former.
The dual local-or-hosted deployment is Exa's sharper edge. Perplexity's server runs as a single package. Exa gives you the choice, which matters if your infrastructure already leans one way on secrets management.
Tool count tells the same story. Exa ships four, split between defaults and opt-ins. Perplexity maps its tools straight to its Agent API presets. Scraping-first MCP servers tend to sprawl into general browsing. These two don't. The surface stays small, search and research only, so an agent rarely fumbles into the wrong tool by accident.
Pricing shapes the choice too, in a quieter way. Both run on credit-metered APIs behind the free tier, so a workload that calls search dozens of times per session will hit a paid plan on either server eventually. Neither MCP layer changes that math. What changes is how visible the cost is to the agent calling the tool, and neither server currently surfaces per-call cost the way some other API-backed MCP servers do.
Gus Marquez