Jina AI Remote MCP Server – Search and Reading Tools

Overview

Jina AI Remote MCP Server is the official hosted Model Context Protocol server for Jina AI. It connects AI agents to Jina Reader, Search, Embeddings, and Reranker capabilities without requiring the user to run Jina infrastructure locally. The server is useful when an assistant needs current web context, clean page extraction, academic search, image discovery, document conversion, or semantic utilities that go beyond the model's built-in knowledge.

What the MCP server enables

The server exposes a broad tool set for information retrieval and content processing. Document and page tools include read_url, parallel_read_url, capture_screenshot_url, guess_datetime_url, and extract_pdf. Search tools include search_web, search_arxiv, search_ssrn, search_images, search_jina_blog, search_bibtex, and parallel search variants. Semantic tools include sort_by_relevance, classify_text, deduplicate_strings, and deduplicate_images.

Several tools require a Jina API key, especially search, query expansion, reranking, classification, deduplication, and PDF extraction. Some reading tools can work without an API key, but Jina documents that unauthenticated or optional-key usage is rate limited and recommends using an API key for higher limits and better performance.

When to use it

Use this MCP server when an agent needs to search the live web, read URLs into clean Markdown, compare multiple sources, find academic papers, generate BibTeX citations, inspect PDFs, rerank candidate documents, classify text into custom labels, or deduplicate text and images. It is especially useful for research assistants, coding agents that need current documentation, and workflows where the model should ground an answer in fresh external pages.

Connection and authentication

The recommended remote endpoint is https://mcp.jina.ai/v1 using Streamable HTTP. Authentication is supplied with an optional Authorization: Bearer ${JINA_API_KEY} header. Clients that do not support remote MCP servers can use the documented mcp-remote proxy through npx, passing the same endpoint and authorization header. Jina also keeps /sse as a backward-compatible endpoint alias, but its documentation states that both /v1 and /sse use Streamable HTTP protocol semantics rather than the deprecated SSE transport.

Key considerations

Registering every available tool can consume context window space because each MCP tool contributes names, descriptions, and schemas. Jina supports server-side filtering with URL query parameters such as include_tools, exclude_tools, include_tags, and exclude_tags so clients can expose only the needed capabilities. Client-side tool caching may require removing and re-adding the server when definitions change. Some clients enforce fixed tool response limits; Jina applies guardrails for large read_url outputs in certain clients to avoid rejected responses. Store API keys securely and avoid hardcoding secrets unless a client cannot expand environment variables.