Exa MCP 🔌 MCP server Freemium

What it does

Exa MCP connects AI agents to Exa, a search engine built for AI. Results come with clean page content, and the agent can read any URL as Markdown. A coding agent can look up current docs, a changelog or a post about an error message, then read the pages instead of guessing. With the optional tools switched on, the agent can also run filtered searches or hand off multi-step research to an Exa Agent. The server is MIT licensed and mainly offered as a hosted endpoint. The repo also ships as an Agent Plugin with two research skills.

Tools it gives your agent

On by default:

• web_search_exa: searches the web and returns clean content
• web_fetch_exa: reads one or more URLs as Markdown

Optional, switched on with the tools URL parameter:

• web_search_advanced_exa: filters by domain and date, and adds highlights, summaries and subpage crawling
• agent_run: runs an Exa Agent for research, list-building and structured output, which requires sign-in

Works with

It is a remote server at https://mcp.exa.ai/mcp. The README covers Claude Code, Codex, Cursor, VS Code, Gemini CLI, OpenCode, Windsurf, Zed and more. Any client that supports remote servers can connect.

How to connect

In Claude Code, install the plugin, which includes the MCP server and skills:

claude plugin install exa@claude-plugins-official

In Codex:

codex mcp add exa --url https://mcp.exa.ai/mcp

To enable more tools, list every tool you want in the URL, because the list replaces the defaults: https://mcp.exa.ai/mcp?tools=web_search_exa,web_fetch_exa,agent_run.

Cost and limits

You can use the hosted server anonymously, with rate limits. If you sign in with OAuth or an API key, you get higher limits and access to Exa Agent. At the time of writing (September 2026), new accounts get $20 in free credits, and the free tier adds $10 every month. After that, search costs $7 per 1,000 requests and page contents cost $1 per 1,000 pages. Agent runs are priced by effort level.

Safety notes

Sign in with OAuth where you can. If you use an API key, send it in the Authorization or x-api-key header rather than in the URL. Treat fetched pages as untrusted input that may carry prompt injection.