Jina AI

The Jina AI Remote MCP Server gives developers direct access to Jina’s powerful suite of AI APIs like Reader API, Embeddings API, and Reranker API.

It handles the heavy lifting of converting web content into clean, structured data that LLMs can actually work with. No more wrestling with messy HTML or getting blocked by anti-scraping measures.

Features

• 🔍 Web Content Extraction – Pull clean markdown content from any URL with read_url, stripping away navigation menus, ads, and other noise that clutters web pages
• 📸 Screenshot Capture – Generate high-quality screenshots of web pages through capture_screenshot_url without setting up your own browser automation
• 🌐 Web Search – Access current information via search_web that returns the top results in LLM-friendly format—no need to parse search engine HTML
• 📚 Academic Research – Find relevant papers using search_arxiv specifically designed for academic content
• 🖼️ Image Search – Locate relevant images across the web with search_images, similar to Google Images but programmatically accessible
• 🔄 Query Expansion – Improve search results with expand_query that rewrites and diversifies your search terms using submodular optimization
• 🎯 Document Reranking – Boost search relevance with sort_by_relevance that applies Jina’s world-class reranking model to your results
• 🧹 Content Deduplication – Eliminate redundant content using deduplicate_strings and deduplicate_images that identify semantically similar items

How To Use It

1. If your development environment supports remote MCP servers (like Cursor or LM Studio), add this configuration:

{
  "mcpServers": {
    "jina-mcp-server": {
      "url": "https://mcp.jina.ai/sse",
      "headers": {
        "Authorization": "Bearer ${JINA_API_KEY}"
      }
    }
  }
}

Replace ${JINA_API_KEY} with your free API key from jina.ai. For tools that don’t require authentication (like basic read_url), you can omit the Authorization header.

2. If your environment doesn’t yet support remote MCP servers, use the mcp-remote proxy:

{
  "mcpServers": {
    "jina-mcp-server": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://mcp.jina.ai/sse",
        "--header",
        "Authorization: Bearer ${JINA_API_KEY}"
      ]
    }
  }
}

This requires Node.js and npm installed on your system.

3. When using the tools, be aware of context window limitations. Some users report getting stuck in tool-calling loops when using thinking models with the default 4096 context window.

The solution? Load your model with a larger context size to accommodate the full tool-calling chain and thought process.

4. For developers who want to customize or self-host, Jina provides a GitHub repository you can clone and deploy:

git clone https://github.com/jina-ai/MCP.git
cd MCP
npm install
npm run start

5. You can also deploy directly to Cloudflare Workers with one click, creating your own instance at jina-mcp-server.<your-account>.workers.dev/sse.

FAQs

Q: Do I need an API key for all features?
A: No. Basic content extraction (read_url) and screenshot capture work without authentication but have lower rate limits. Web search, arXiv search, image search, query expansion, and deduplication tools require a Jina API key. You can get a free key with generous usage limits from their website.

Q: How does the deduplication actually work?
A: The deduplicate_strings and deduplicate_images tools use Jina’s Embeddings API to create vector representations of your content, then apply submodular optimization to select the most diverse yet relevant items. It’s smarter than simple string matching—it identifies semantically similar content even when wording differs.

Q: I’m hitting rate limits with the free tier. What are my options?
A: Free API keys come with 10 million tokens. If you exceed this, you can either wait for the next billing period (limits reset monthly) or upgrade to a premium key for higher limits. For production applications with heavy usage, contact Jina about enterprise plans.

Q: Why do I get stuck in tool-calling loops sometimes?
A: This happens when your model’s context window fills up during extended tool usage. Thinking models like gpt-oss-120b generate many intermediate steps. Increase your context window size—4096 is often too small for complex tool chains. Set it to at least 8K or 16K for smoother operation.

Q: How does this compare to building my own scraping solution?
A: You’d need to handle browser automation, proxy rotation, anti-bot bypass, HTML parsing, and content extraction—plus ongoing maintenance as websites change. This server handles all that infrastructure for you, with better results and less development time. The token cost often works out cheaper than the engineering hours required.