Exa

Exa built its search index specifically for retrieval by language models rather than adapting a traditional keyword search engine, and this MCP server exposes that as tools for web search, page-content fetching, and chained multi-step research rather than a single query-and-summarize call. The distinction shows up most on research tasks that need several rounds of searching and reading before an answer is assembled, rather than a one-shot lookup.

Two run modes are available: connect directly to Exa’s hosted endpoint at https://mcp.exa.ai/mcp, which handles authentication via OAuth with most clients prompting for it automatically, or run the exa-mcp-server npm package locally via npx against an EXA_API_KEY obtained from the Exa dashboard. The hosted option needs no local process at all, which makes it a low-friction default when a client supports it.

Client support is unusually wide for an MCP server — beyond the standard set, it’s documented working with Codex, Kiro, LM Studio, Replit, Gemini CLI, Zed, Warp, and v0 by Vercel, among others — reflecting Exa’s positioning as infrastructure meant to plug into whatever agent tooling a team is already using.