Exa MCP Server

Latest Audit Result

exa-js is pinned at ^2.8.0 (lockfile 2.8.0) vs latest 2.12.1 → outdated by 4 minor releases. The MCP server does NOT use any high-level exa-js methods (search/searchAndContents/getContents/findSimilar/answer/research); exa-js is used purely as a thin HTTP wrapper via exa.request(). Implemented Exa endpoints: /search (web_search_exa, web_search_advanced_exa, get_code_context_exa, deep_search_exa, company_research_exa, people_search_exa, linkedin_search_exa), /contents (web_fetch_exa, crawling_exa alias), and /research/v1 + /research/v1/{id} (deep_researcher_start/check). NOT implemented: /findSimilar and /answer (no MCP tool, no SDK call). /context is declared as a constant but never called anywhere. Search types exposed are 'auto'|'fast'|'instant'; 'neural' and 'keyword' are not exposed. Content options used: text, highlights, summary, subpages, livecrawl (hard-coded to 'fallback'; the full mode enum is not user-selectable), maxAgeHours, livecrawlTimeout. Most legacy tools (company_research, people_search, linkedin_search, deep_search, get_code_context, deep_researcher_*, crawling_exa) are flagged Deprecated in the registry; the canonical surface is web_search_exa + web_search_advanced_exa + web_fetch_exa.

Missing Features (7)

• find_similar (no findSimilar / /findSimilar usage; no MCP tool exposed)
• answer (no exa.answer / exa.streamAnswer / /answer usage; no MCP tool exposed)
• neural search type (only 'auto'|'fast'|'instant' are exposed in tool schemas)
• keyword search type (not exposed)
• /context endpoint declared in API_CONFIG.ENDPOINTS.CONTEXT but never called
• explicit livecrawl mode param ('always'|'never'|'preferred' supported in types but only 'fallback' hard-coded)
• exa-js high-level SDK methods unused — server uses SDK only as a thin HTTP client via exa.request()