出海助手-Exa联网搜索 / Exa MCP Search

> 本技能改造自开源技能 exa-search（来源：mxyhi/ok-skills，ok-skills 精选 Agent 技能库）。已完整保留原始运行文件（scripts/references/assets 与 LICENSE），并新增中文出海使用说明。原始英文说明见文末「English (Original)」与同目录 SKILL.original.md。

功能说明 / Functional Overview

通过 Exa 语义搜索引擎获取高质量外网信息，作为 Tavily 之外的备选。

适用场景 / Use Cases

• 海外市场/技术检索
• 语义化精准搜索
• 竞品信息搜集

使用方法 / How to Use

1. 配置 Exa API Key 后，用自然语言描述检索需求即可。

示例 / Examples

• 「海外市场/技术检索」——用自然语言把任务交给本技能即可。

限制说明 / Limitations

• 需要外部服务/凭证的技能，请先完成授权配置后使用。
• 本技能为开源技能的中文本地化版本，核心逻辑与原技能一致；如遇专业术语，正文保留英文原词（如 Shopify、SEO、GA4 等）。

English (Original) / 英文原文（完整保留）

> The section below is the original skill definition, preserved verbatim for traceability and full functionality.

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