The search takes the page you want in words and hands back the text rather than a list of links to open, which removes the fetch-and-strip step that usually follows. When the returned content is not enough, the fetch call reads whole pages — several URLs at once.
Two calls by default: a web search that returns clean text from the top results, and a fetch that reads one URL or several as markdown. Two more — an advanced search with filters and an Exa Agent run — switch on when you ask for them.
- A web search that takes a description of the ideal page — "blog post comparing React and Vue performance" rather than a keyword string — and returns clean, ready-to-use content from the top results (`web_search_exa`)
- Any URL read in full as clean markdown, several at a time (`web_fetch_exa`)
- An advanced search with filters, domains, dates, highlights, summaries and subpage crawling, when the plain search is too blunt (`web_search_advanced_exa`)
- A multi-step Exa Agent run for research, list-building, enrichment and structured output (`agent_run`)
- Two agent skills that add orchestration on top — one for deep research and competitive analysis, one for Exa Agent runs and enrichment
Nothing to install: the hosted server at `https://mcp.exa.ai/mcp` works anonymously, rate-limited. Signing in over OAuth raises the limits and unlocks Exa Agent — `https://mcp.exa.ai/mcp?login` forces the flow when a client will not prompt — or pass an API key as an `Authorization: Bearer` or `x-api-key` header. The optional tools are switched on by appending `?tools=` to the URL, and that list replaces the defaults rather than adding to them, so name every tool you want. A local `exa-mcp-server` package over stdio is also published.
One command — npx -y exa-mcp-server