An Exa MCP server gives an AI client neural (semantic) web search plus page crawling — great for research where keyword search falls short. Exa offers both a hosted remote endpoint and a local npx server. Here's the setup.
Option A — hosted remote (no key in config)
The quickest route bridges to Exa's hosted MCP via mcp-remote:
{
"mcpServers": {
"exa": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.exa.ai/mcp"]
}
}
}
You can also pass the key in the URL: https://mcp.exa.ai/mcp?exaApiKey=YOUR_KEY.
Option B — local npx with API key
Grab a key from dashboard.exa.ai/api-keys, then:
{
"mcpServers": {
"exa": {
"command": "npx",
"args": ["-y", "exa-mcp-server"],
"env": {
"EXA_API_KEY": "your_api_key"
}
}
}
}
Claude Code one-liner:
claude mcp add --transport stdio --env EXA_API_KEY=your_key exa -- npx -y exa-mcp-server
Restart the client; see MCP config file location.
Verify
Ask: "Use Exa to find recent technical write-ups about MCP security and list the top 5 with links." Relevant results back confirms it works.
Common problems
401/ quota — the key is wrong or out of credits; check the dashboard.- Remote bridge fails —
mcp-remoteneeds Node 18+ and outbound network; see Cursor MCP not working. - Thin results — neural search rewards descriptive queries; phrase the ask like you'd describe the page.
Security
Search servers fetch and return external content, which is a prompt-injection surface — treat retrieved text as untrusted. See MCP prompt injection prevention and MCP security best practices.
Exa vs the alternatives
Exa leans neural/semantic; Tavily is tuned for agent RAG. Compare Tavily MCP vs Exa MCP, or for scraping see Firecrawl. See best MCP for browser automation.
Going further
See the Exa agent profile, browse the search category, or grab a content-writer loadout.