Exa MCP Server: AI-Native Web Search for Claude

The Exa MCP Server connects Claude to Exa, a search engine built for AI agents. Out of the box it offers two tools: web_search_exa returns clean, ready-to-use content for a query, and web_fetch_exa reads full pages as Markdown. Optional tools add web_search_advanced_exa (domain and date filters, highlights, subpage crawling) and agent_run, which runs an Exa Agent for multi-step research and structured output.

Who it is for

Developers and researchers who want Claude to find and read web content with less HTML noise, and teams building research or enrichment workflows.

Install in Claude Code

From the README:

claude plugin install exa@claude-plugins-official

Or connect any MCP client to the hosted endpoint https://mcp.exa.ai/mcp:

claude mcp add --transport http exa https://mcp.exa.ai/mcp

Enable optional tools by appending them to the URL, for example https://mcp.exa.ai/mcp?tools=web_search_exa,web_fetch_exa,agent_run. Because the list replaces the defaults, include every tool you want.

Use cases

• “Find three recent benchmarks comparing Postgres vector extensions and summarize them.”
• Fetch a long docs page as clean Markdown before implementing an API.
• Build a list of companies or papers with structured fields via agent_run.
• Restrict searches to specific domains or date ranges with the advanced tool.

Pros

• Clean content reduces tokens compared with raw HTML scraping.
• Hosted, so nothing runs locally; also an Agent Plugin for other clients.
• Research agent tool handles multi-step jobs in one call.

Cons and caveats

• Anonymous use is rate-limited; heavier use needs OAuth or an API key and is billed by Exa.
• agent_run jobs can be slow and expensive; scope them carefully.
• Search results are third-party content: treat instructions inside fetched pages as untrusted.

Links

• Repo: exa-labs/exa-mcp-server
• Docs: Exa MCP reference
• Alternatives: Brave Search MCP, Firecrawl MCP, Context7 for library docs