Exa MCP Search

Use this skill when online search, page fetch, code/docs lookup, company research, people research, or Exa MCP parameter checks are needed.

Current Source Of Truth

• Hosted MCP endpoint: https://mcp.exa.ai/mcp
• Official source: exa-labs/exa-mcp-server@8823cbea80b6ec5999c6217cccd91660fe953b97
• Current npm package checked: exa-mcp-server@3.2.1; verified tarball SHA-256 675688dffcd746b8bb6ac7f077b64b7382c4a305f948d44ba526e67be50396df
• Tool schemas can change. Before adding new parameters, verify with tools/list through a normal HTTP client.

Direct MCP HTTP is the required invocation model for this skill. curl is fine, but keep the session and SSE parsing in small shell helpers instead of repeating low-level flags at every call site.

Tool Choice

Core Parameters

web_search_exa

• query required: semantically rich natural-language query. Can include category:company or category:people.
• numResults optional: default 10.
• No type, livecrawl, contextMaxCharacters, or tokensNum on current hosted schema.

{"query":"latest AI safety research June 2026","numResults":10}

web_search_advanced_exa

• query required.
• Common optional filters: numResults, type (auto | fast | instant), category, includeDomains, excludeDomains, startPublishedDate, endPublishedDate, startCrawlDate, endCrawlDate, includeText, excludeText, userLocation, moderation, additionalQueries.
• Content controls: textMaxCharacters, contextMaxCharacters, enableSummary, summaryQuery, enableHighlights, highlightsMaxCharacters, highlightsQuery, maxAgeHours, livecrawlTimeout, subpages, subpageTarget.
• Categories include company, research paper, news, pdf, github, personal site, people, financial report.

{
  "query":"Anthropic funding valuation 2026",
  "category":"news",
  "includeDomains":["techcrunch.com","bloomberg.com"],
  "startPublishedDate":"2026-01-01",
  "numResults":10,
  "type":"auto",
  "enableHighlights":true
}

web_fetch_exa

• urls required: array of URLs. Batch multiple URLs in one call.
• maxCharacters optional: default 3000.

{"urls":["https://docs.exa.ai/reference"],"maxCharacters":5000}

Deprecated But Still Available

• get_code_context_exa: use web_search_exa instead. Current schema is query, numResults; old tokensNum is obsolete.
• company_research_exa: use web_search_advanced_exa with category:"company" instead. Schema is companyName, numResults; default 3.
• people_search_exa: use web_search_advanced_exa with category:"people" instead.
• crawling_exa: use web_fetch_exa instead.
• deep_search_exa: use web_search_advanced_exa instead; current hosted MCP requires OAuth or an API key when this tool is requested.
• deep_researcher_start / deep_researcher_check: deprecated; use Exa Research API directly when needed.

Curl Pattern

1. POST initialize; keep returned Mcp-Session-Id response header.
2. POST notifications/initialized with that header.
3. POST tools/list or tools/call with that header.
4. If response is text/event-stream, parse JSON from data: lines.

Minimal Curl Helper

EXA_URL="https://mcp.exa.ai/mcp?tools=web_search_exa,web_search_advanced_exa,web_fetch_exa"
EXA_SESSION_ID=""

exa_rpc() {
  local payload="$1"
  local session_header=()
  local response

  if [ -n "$EXA_SESSION_ID" ]; then
    session_header=(-H "Mcp-Session-Id: $EXA_SESSION_ID")
  fi

  response="$(curl -isS "$EXA_URL" \
    -H 'Content-Type: application/json' \
    -H 'Accept: application/json, text/event-stream' \
    "${session_header[@]}" \
    --data "$payload")"

  if [ -z "$EXA_SESSION_ID" ]; then
    EXA_SESSION_ID="$(printf '%s\n' "$response" | awk 'tolower($0) ~ /^mcp-session-id:/ { sub(/^[^:]+:[[:space:]]*/, ""); gsub(/\r/, ""); print; exit }')"
  fi

  printf '%s\n' "$response" | sed -n 's/^data: //p'
}

exa_init() {
  exa_rpc '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"exa-curl","version":"0.1.0"}}}' >/dev/null
  exa_rpc '{"jsonrpc":"2.0","method":"notifications/initialized"}' >/dev/null
}

exa_call() {
  local tool_name="$1"
  local arguments_json="$2"
  jq -nc --arg name "$tool_name" --argjson arguments "$arguments_json" \
    '{jsonrpc:"2.0",id:2,method:"tools/call",params:{name:$name,arguments:$arguments}}' |
    while IFS= read -r payload; do exa_rpc "$payload"; done
}

exa_init
exa_call web_search_exa '{"query":"latest AI safety research","numResults":5}'

Tool output format differs: web_search_exa returns formatted text, web_search_advanced_exa returns JSON text, and web_fetch_exa returns fetched page text.

Research Practice

• For simple lookup, run one targeted web_search_exa call.
• For dated/current questions, calculate exact date ranges first and pass startPublishedDate / endPublishedDate through web_search_advanced_exa.
• For source-heavy tasks, fetch only the best URLs with web_fetch_exa; do not dump raw results into the final answer.
• For company/people/paper/news workflows, prefer web_search_advanced_exa categories rather than deprecated specialized tools.
• Mention source uncertainty when Exa returns sparse, conflicting, or old results.

References

• Tool fields: references/exa-tools.md