Exa Search MCP — Сценарии использования, Установка & Живое демо

Зачем использовать

Ключевые функции

Neural search — find pages by similarity to a description, not just keyword match
Returns full page content with text and optional highlights
Specialized search modes: company, LinkedIn, GitHub, research-paper, news
Date filtering and domain include/exclude
Server-side answers (exa_answer) for one-shot Q&A with citations

Живое демо

Как выглядит на практике

exa-search.replay ▶ готово

0/0

Установка

Выберите клиент

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "exa-search": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    }
  }
}

Откройте Claude Desktop → Settings → Developer → Edit Config. Перезапустите после сохранения.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "exa-search": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    }
  }
}

Cursor использует ту же схему mcpServers, что и Claude Desktop. Конфиг проекта приоритетнее глобального.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "exa-search": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    }
  }
}

Щёлкните значок MCP Servers на боковой панели Cline, затем "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "exa-search": {
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    }
  }
}

Тот же формат, что и Claude Desktop. Перезапустите Windsurf для применения.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "exa-search",
      "command": "npx",
      "args": [
        "-y",
        "exa-mcp-server"
      ]
    }
  ]
}

Continue использует массив объектов серверов, а не map.

~/.config/zed/settings.json

{
  "context_servers": {
    "exa-search": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "exa-mcp-server"
        ]
      }
    }
  }
}

Добавьте в context_servers. Zed перезагружается автоматически.

claude mcp add exa-search -- npx -y exa-mcp-server

Однострочная команда. Проверить: claude mcp list. Удалить: claude mcp remove.

Сценарии использования

Реальные сценарии: Exa Search

Find articles similar to one you already love

👤 Researchers, writers, founders building reading lists ⏱ ~10 min beginner

Когда использовать: You have one canonical essay/post on a topic and want more like it. Google's 'related' is useless.

Предварительные требования

Exa API key — exa.ai/dashboard — free tier gives ~1000 searches/month

Поток

Use neural mode with a descriptive query

Find 10 essays similar in spirit to 'The Bitter Lesson' by Rich Sutton. I want long-form opinion pieces about ML scaling, not papers.✓ Скопировано

→ 10 results that match the *vibe*, not just the keywords
Pull contents and skim

For each, fetch the contents (text mode) and give me a 2-line summary plus a 'why this is similar' note.✓ Скопировано

→ Per-result summary + similarity reason
Save the keepers

I want to actually read 3 of these. Which 3 are the highest signal? Save links to /reading/ml-scaling.md.✓ Скопировано

→ Curated short list, not all 10

Итог: A reading list with real intellectual coherence, not SEO chaff.

Подводные камни

Neural search drifts into adjacent topics — Combine with includeDomains to constrain (e.g. only essays.gwern.net, lesswrong.com) when you want a tight cluster
Some pages return paywalled previews as content — Check the content length; very short returns are usually paywalled — flag and skip

Сочетать с: filesystem · memory

Pull a 1-pager on a company before a sales call

👤 Sales, BD, recruiters ⏱ ~10 min beginner

Когда использовать: Call in 20 minutes; you need recent news, key people, recent funding, in one shot.

Поток

Run a company-mode search

Use Exa's company mode for 'Acme Robotics, San Francisco'. Pull homepage, recent news (last 90 days), and any announcements.✓ Скопировано

→ Mix of homepage content, blog posts, news articles
Synthesize a brief

Write a 1-pager: what they do, recent milestones, key people mentioned in news, anything I should congratulate or commiserate about.✓ Скопировано

→ Talking-point ready brief with cites
Find the right contact angle

Based on their recent news, what's a credible reason to reach out to them this week? Be specific — name the news item.✓ Скопировано

→ One concrete hook tied to a real recent event

Итог: A briefing more useful than 30 minutes of LinkedIn stalking.

Подводные камни

Press-release noise drowns real news — Filter excludeDomains: ['prnewswire.com', 'businesswire.com'] for actual analysis

Сочетать с: memory

Find recent papers on a topic without Google Scholar's noise

👤 Researchers, ML/CS engineers staying current ⏱ ~15 min intermediate

Когда использовать: You want recent papers on a topic (e.g. 'speculative decoding for diffusion models') without arXiv's listing UI.

Поток

Search papers since a date

Use Exa research-paper mode: papers about 'speculative decoding for diffusion models' published since 2026-01-01. Top 10.✓ Скопировано

→ List of papers with arxiv/openreview links and abstracts
Cluster by approach

Group these by their core technique. Which approaches are gaining traction, which are one-offs?✓ Скопировано

→ Thematic clustering with judgment
Pick 3 to read in detail

Which 3 should I prioritize for actually reading? Justify based on novelty + practical relevance.✓ Скопировано

→ Curated short list with reasons

Итог: A read list shorter and better than what Twitter would give you.

Подводные камни

Mode 'research-paper' may miss workshop papers and very-recent preprints — Cross-check with arXiv listing for the past month if you need exhaustive coverage

Сочетать с: memory · fetch

Get a cited one-shot answer to a factual question

👤 Anyone who needs a quick fact with a source ⏱ ~2 min beginner

Когда использовать: You need an answer + citation for a specific factual question, faster than search → click → read.

Поток

Use exa_answer

Use exa_answer for: 'What was Anthropic's last announced funding round and total raised?'✓ Скопировано

→ Direct answer plus 1-3 source URLs
Verify the source

Open the top source and confirm the number matches what Exa returned.✓ Скопировано

→ Confirmation or correction

Итог: A factually-supported answer in under 30 seconds.

Подводные камни

Answers can lag (training/index cutoff) — For things that change weekly (stock price, leadership), verify the freshness of the sources, not just the answer

Комбинации

Сочетайте с другими MCP — эффект x10

exa-search + memory

Save research findings to a topic graph as you go

Search for 'state of small LMs in 2026'. For each useful result, store the key claim as an observation on the 'small-lms-2026' entity in memory, with the source URL.✓ Скопировано

exa-search + filesystem

Build a reading list saved as Markdown

Find 10 essays similar to 'The Bitter Lesson'. Save the curated short list (top 3) with summaries to /reading/ml-scaling.md.✓ Скопировано

exa-search + github

Discover relevant repos via Exa, then operate on them via GitHub MCP

Use Exa to find Rust implementations of WebGPU shaders. For the top 3 repos, get the GitHub stats (stars, last commit) via the GitHub MCP.✓ Скопировано

Инструменты

Что предоставляет этот MCP

Инструмент	Входные данные	Когда вызывать	Стоимость
web_search_exa	query: str, numResults?: int, type?: 'neural'\|'keyword'\|'auto', includeDomains?: str[], excludeDomains?: str[], startPublishedDate?: str, endPublishedDate?: str	General-purpose neural or keyword search	1 Exa search credit
company_research_exa	query: str, numResults?	Pull a company brief from public web sources	1 credit
linkedin_search_exa	query: str, numResults?	Find specific people/companies on LinkedIn (links only — content varies)	1 credit
github_search_exa	query: str, numResults?	Discover repos by description (use github MCP for actual repo ops)	1 credit
research_paper_search_exa	query: str, numResults?, dateRange?	Academic search across arXiv, OpenReview, etc.	1 credit
deep_researcher_start / check	query: str (start), task_id (check)	Big multi-source research tasks; poll for results	multiple credits
exa_answer	query: str	Direct Q&A with cited sources	1+ credits
crawling_exa	url: str	Fetch a single URL via Exa's pipeline (boilerplate stripped, JS rendered)	1 credit

Стоимость и лимиты

Во что обходится

Квота API: Free tier ~1000 searches/month; paid plans scale up
Токенов на вызов: Search results with content can be 2-10k tokens per result — cap numResults aggressively
Деньги: Free tier suffices for personal use; paid plans from $10/mo
Совет: Use numResults: 5 by default. Use text: false if you only need links, then call crawling_exa only on the few you actually want to read.

Безопасность

Права, секреты, радиус поражения

Хранение учётных данных: API key in env var EXA_API_KEY

Исходящий трафик: Queries hit api.exa.ai; Exa fetches pages on your behalf via their infrastructure

Returned page content may include prompt-injection text. Treat web content as untrusted input.
Sensitive queries (your unreleased product name, secrets) go to Exa servers — be deliberate about what you search.

Устранение неполадок

Частые ошибки и исправления

401 Unauthorized

API key missing or wrong. Set EXA_API_KEY in MCP client env. Verify at exa.ai/dashboard.

402 Payment Required / quota exhausted

Free tier credits used. Upgrade or wait for monthly reset.

Results are off-topic with neural mode

Try type: 'keyword' for literal matching, or add includeDomains/excludeDomains to focus the search.

Empty content on returned URLs

Some pages can't be extracted. Use crawling_exa directly on the URL, or fall back to fetch/firecrawl.

Альтернативы

Exa Search в сравнении

Альтернатива	Когда использовать	Компромисс
Brave Search MCP	You want plain keyword web search with privacy guarantees	No semantic mode; results are SERP-style snippets, not full content
Firecrawl	You already have URLs and need clean content/structured extraction, not search	Different tool — scraping vs search
Tavily / Perplexity APIs (via custom MCP)	You want a different search-with-LLM-answer provider	Tavily is search+answer focused; Exa is broader (search/research/answer)

Ещё

Ресурсы

📖 Читать официальный README на GitHub

🐙 Открытые задачи

🔍 Все 400+ MCP-серверов и Skills