jina-ai/MCP

What it solves

This project provides a remote Model Context Protocol (MCP) server that allows LLMs to access a suite of web-based tools for content extraction, web search, and semantic analysis. It solves the problem of LLMs having outdated or limited knowledge by giving them the ability to read web pages, search the web, and process academic papers from arXiv and SSRN, all through a standardized protocol.

How it works

The server implements the Model Context Protocol (MCP) using Streamable HTTP transport. It acts as a bridge between an MCP-compatible client (like Cursor, Claude Desktop, or LM Studio) and Jina AI's APIs (Reader, Embeddings, and Reranker). When an LLM requests a tool, the server calls the corresponding Jina AI API to perform the tasks—such as converting a URL to markdown, performing a deep web search, and reranking documents for relevance.

Who it’s for

Developers and users of AI agents and LLM clients that support the MCP standard, seeking to integrate high-quality web search, academic research, and web content extraction into their AI's workflow.

Highlights

• Comprehensive Toolset: Includes 22 tools for web search (web, arXiv, arXiv, SSRN), screenshot capture, and PDF extraction (figures, tables, equations).
• Deep Search: The search_web_deep tool reads each result page and uses a Reranker API to return the best paragraph-length passages from each page.
• Context Window Optimization: Supports server-side filtering via query parameters (include_tools, exclude_tools) to prevent LLMs from wasting context tokens on unused tool definitions.
• Parallelism: Offers parallel_* versions of search and read tools to execute multiple queries or URLs concurrently for efficiency.
• Question-Grounded Reading: The read_url tool can take a question parameter to return only the most relevant ranked passages from a page instead of the full content.
• Capped Responses: Implements a token guardrail to prevent responses from being rejected by clients with fixed token limits (e.g., 25k tokens).