Jina AI MCP — Anwendungsfälle, Installieren & Live-Demo

Warum nutzen

Hauptfunktionen

First-party — official Jina AI MCP
read_url returns clean markdown — handles JS-rendered sites
Search across web, arXiv, SSRN, images, and BibTeX in one interface
Processing tools: rerank, classify, dedupe (text and images), extract_pdf

Live-Demo

In der Praxis

jina.replay ▶ bereit

0/0

Installieren

Wählen Sie Ihren Client

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

{
  "mcpServers": {
    "jina": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.jina.ai/sse"
      ]
    }
  }
}

Öffne Claude Desktop → Settings → Developer → Edit Config. Nach dem Speichern neu starten.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "jina": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.jina.ai/sse"
      ]
    }
  }
}

Cursor nutzt das gleiche mcpServers-Schema wie Claude Desktop. Projektkonfiguration schlägt die globale.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "jina": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.jina.ai/sse"
      ]
    }
  }
}

Klicken Sie auf das MCP-Servers-Symbol in der Cline-Seitenleiste, dann "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "jina": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.jina.ai/sse"
      ]
    }
  }
}

Gleiche Struktur wie Claude Desktop. Windsurf neu starten zum Übernehmen.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "jina",
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.jina.ai/sse"
      ]
    }
  ]
}

Continue nutzt ein Array von Serverobjekten statt einer Map.

~/.config/zed/settings.json

{
  "context_servers": {
    "jina": {
      "command": {
        "path": "npx",
        "args": [
          "-y",
          "mcp-remote",
          "https://mcp.jina.ai/sse"
        ]
      }
    }
  }
}

In context_servers hinzufügen. Zed lädt beim Speichern neu.

claude mcp add jina -- npx -y mcp-remote https://mcp.jina.ai/sse

Einzeiler. Prüfen mit claude mcp list. Entfernen mit claude mcp remove.

Anwendungsfälle

Praxisnahe Nutzung: Jina AI

Digest recent arXiv papers on a topic

👤 Researchers, ML engineers staying current ⏱ ~20 min intermediate

Wann einsetzen: You want to know what's new on arXiv about your topic without reading 50 abstracts.

Voraussetzungen

Optional Jina API key — jina.ai → dashboard → API key (free tier works for light use)

Ablauf

Search arXiv

Use search_arxiv to find papers from the last 30 days about 'speculative decoding for LLM inference'. Return top 20.✓ Kopiert

→ Paper list with titles, authors, abstracts
Rerank by relevance

Use sort_by_relevance to rerank against this query: 'practical speedups in production inference, not pure research'. Keep top 8.✓ Kopiert

→ Reranked list
Summarize each

For the top 8, extract_pdf the paper, summarize in 3 bullets: contribution, method, reported speedup. Output as a markdown table.✓ Kopiert

→ Digest-ready summary table

Ergebnis: A weekly research digest on your topic in 10 minutes.

Fallstricke

extract_pdf on every result is expensive — credits add up — Rerank first to cut candidates, only extract the top N

Kombinieren mit: notion

Convert a batch of URLs into clean markdown for RAG

👤 AI engineers building retrieval systems ⏱ ~15 min intermediate

Wann einsetzen: You have a list of URLs to ingest. You want clean markdown, not raw HTML or a parsing pipeline.

Ablauf

Read URLs in parallel

Use parallel_read_url on this list [URLs]. Return markdown for each with original URL as key.✓ Kopiert

→ Markdown per URL
Dedupe near-duplicates

Use deduplicate_strings at 0.9 similarity to drop near-duplicate pages (common for mirror docs).✓ Kopiert

→ Deduped set with IDs of dropped pages
Save to disk

Save each to ./knowledge/<slug>.md where slug is derived from the URL path.✓ Kopiert

→ Markdown files ready for embedding pipeline

Ergebnis: A clean corpus for your embedding/indexing step, without writing any scraping code.

Fallstricke

Paywalled or JS-auth-walled pages return blank/garbage — Spot check a few URLs — if the content is thin, fall back to playwright for auth flows

Kombinieren mit: filesystem · firecrawl

Classify a batch of text with custom labels

👤 Data analysts, growth teams ⏱ ~15 min beginner

Wann einsetzen: You have N free-text items (tickets, reviews, survey responses) and want them bucketed into your taxonomy.

Ablauf

Define labels

My labels: ['bug', 'feature_request', 'question', 'praise', 'other']. Sample the first 10 items and sanity-check the labels fit.✓ Kopiert

→ Labels validated against samples
Batch classify

Use classify_text on all items with those labels. Return {id, text, label, confidence}.✓ Kopiert

→ Labelled dataset
Review low-confidence

Flag items where confidence < 0.6 for manual review. Summarize: distribution, outliers, likely missing labels.✓ Kopiert

→ Review queue + taxonomy feedback

Ergebnis: A labeled dataset without fine-tuning a classifier or writing prompts per item.

Fallstricke

Labels are ambiguous and classifier flip-flops on near-ties — Make labels mutually exclusive; if items span categories, allow multi-label output

Kombinieren mit: filesystem

Kombinationen

Mit anderen MCPs für 10-fache Wirkung

jina + notion

Weekly research digest posted to Notion

Search arXiv for new 'agentic RAG' papers this week. Summarize each and create a Notion page in the Research Digest database.✓ Kopiert

jina + firecrawl

Jina for single pages, Firecrawl for full crawls — same clean-markdown output

For the list of URLs, use parallel_read_url (Jina). For the 3 full docs sites, use Firecrawl crawl. Merge into one knowledge dir.✓ Kopiert

jina + filesystem

Build a local markdown knowledge base from a reading list

Read each URL in urls.txt, dedupe, save to ./knowledge/<hash>.md. Overwrite only if content changed.✓ Kopiert

Werkzeuge

Was dieses MCP bereitstellt

Werkzeug	Eingaben	Wann aufrufen	Kosten
search_web	query, num_results?	General web search	credits per call
search_arxiv / search_ssrn / search_bibtex / search_images / search_jina_blog	query	Targeted searches	credits per call
parallel_search_web / parallel_search_arxiv / parallel_search_ssrn	query[]	Multi-query research in one call	credits × N queries
read_url	url	Clean content extraction from any URL	credits per page
parallel_read_url	url[]	Batch webpage ingestion	credits × N pages
capture_screenshot_url	url	Visual snapshot of a page	credits
sort_by_relevance	documents, query	Rerank after search for quality	credits
classify_text	texts, labels	Zero-shot classification	credits per text
deduplicate_strings / deduplicate_images	items, threshold	Remove near-duplicates from a corpus	credits
extract_pdf	url or file	Get structured content from PDFs	credits per PDF
expand_query / primer / guess_datetime_url	utility	Helpers around search tuning	credits (minor)

Kosten & Limits

Was der Betrieb kostet

API-Kontingent: Free tier available with rate limits; paid tiers scale
Tokens pro Aufruf: Output is the bigger cost — PDFs and dedupes can return 10k+ tokens
Kosten in €: Jina API credits, typically measured per-request. See jina.ai/pricing.
Tipp: Rerank before extracting — extract_pdf is expensive. Cache read_url outputs locally; most pages don't change daily.

Sicherheit

Rechte, Secrets, Reichweite

Credential-Speicherung: JINA_API_KEY env var (optional for many tools, required for heavy use)

Datenabfluss: All calls to api.jina.ai / r.jina.ai / s.jina.ai — queries and URLs visible to Jina

Fehlerbehebung

Häufige Fehler und Lösungen

429 Too Many Requests

Free tier has low rate limits. Add a JINA_API_KEY env var and upgrade at jina.ai for burst capacity.

read_url returns empty markdown

Page may be auth-walled or bot-blocked. Try with different User-Agent via tool options, or fall back to playwright/firecrawl.

classify_text assigns everything to 'other'

Your labels may be too narrow or too similar. Add label descriptions ('bug: user reports something broken') for better zero-shot accuracy.

search_arxiv misses recent papers

arXiv index may lag; cross-check with a direct arxiv.org search. Use expand_query to broaden terms.

Alternativen

Jina AI vs. andere

Alternative	Wann stattdessen	Kompromiss
Firecrawl	You need full-site crawls or JSON-schema extraction	Crawl-focused; Jina's superpower is the breadth of processing tools beyond just reading
Exa Search MCP	You want semantic/neural web search as a primary interface	Stronger on semantic retrieval; narrower than Jina's toolbox
Brave Search MCP	You want independent search index + privacy	Search only, no reader/rerank/classify

Mehr

Ressourcen

📖 Offizielle README auf GitHub lesen

🐙 Offene Issues ansehen

🔍 Alle 400+ MCP-Server und Skills durchsuchen