Exa MCP Server

Exa approached search from the opposite direction to Google: instead of ranking pages for humans to click, it built a neural index designed for machines to consume. Queries match on meaning rather than keywords, and results come back with actual page content — not just links — formatted for a model to use immediately. Its MCP server is the cleanest way into that index.

The semantic matching changes what you can ask. "Startups building open-source alternatives to Datadog" is a terrible keyword query and a great Exa query — it describes the kind of thing you want, and the index finds pages that are that kind of thing. This makes Exa disproportionately good for research tasks: competitive landscapes, finding companies or people matching a profile, chasing down writing on a niche topic.

The server ships focused variants beyond plain web_search_exa: company research, LinkedIn search, crawling a specific URL, and a deep researcher mode where an Exa-side agent researches a question across many sources and returns a structured report. For developers, Exa's separate exa-code context tool searches billions of GitHub repos, docs pages and Stack Overflow posts to ground coding answers — a distinct angle on the anti-hallucination problem Context7 attacks via official docs.

It's remote-first (mcp.exa.ai) with an npm local option, and free-tier keys make it cheap to trial. Sensible mental model: Brave for "give me today's links", Exa for "find me the right sources and hand them to my model pre-chewed."

Where it shines in practice

• Phrase queries as a description of the page you want, not as keywords: "blog post by an engineer who migrated a team off Kubernetes" beats "kubernetes migration blog" by a wide margin, because the index is matching intent rather than tokens.
• Keep result counts small and let highlights rather than full text carry the payload — pulling complete page bodies for twenty results is the fastest way to spend a context window on boilerplate.
• Domain include/exclude lists turn a broad neural search into a scoped one, which is how you keep a competitive scan away from listicle farms and content aggregators.
• The deep researcher is genuinely asynchronous: deep_researcher_start hands back a task ID, and the answer arrives via check minutes later, so don't design a flow that expects it in the same turn.