arxiv-reader-mcp
Want to search arXiv papers, fetch metadata, and extract full-text PDFs without leaving your editor? This MCP server connects any MCP-compatible client (Claude Code, etc.) directly to arXiv.
A local server that lets AI assistants like Claude search, browse, and read full papers from arXiv, the giant open-access library of science and computer science research, without leaving your editor or chat window.
arxiv-reader-mcp is a server that connects AI coding assistants and chat tools to arXiv, the large open-access repository of scientific papers used widely in physics, mathematics, computer science, and related fields. The server uses the Model Context Protocol (MCP), which is a standard that lets AI tools call external services in a structured way. Once installed, an AI assistant can search for papers, pull metadata, and even read the full text of a PDF, all without leaving the editor or conversation window.
The server exposes five tools. One searches by keyword, author, category, or date range. A second does a quick keyword search with optional filters. A third fetches the title, authors, abstract, and PDF link for a specific paper ID. A fourth returns recent papers in a chosen subject category. The fifth downloads and extracts the full text from a paper's PDF, which lets an AI assistant read and summarize the actual content rather than just the abstract.
arXiv organizes papers by category codes. The README lists common ones: cs.AI for artificial intelligence, cs.LG for machine learning, cs.CV for computer vision, cs.CL for computation and language, and many others across physics, statistics, and biology. Users can filter any search to a specific category.
Setting it up requires Python and a tool called uv for running Python projects. The server runs as a local process that communicates with a compatible client over standard input and output. To connect it to Claude on claude.ai, you add a short JSON block to the MCP settings file pointing at the server's run command. The README also mentions an MCP Inspector tool for testing the server interactively before connecting a full client.
The project is open source under the MIT license and includes example prompts showing what you can ask an AI once the server is connected: finding recent papers on a topic, summarizing a specific paper by its arXiv ID, or reading the methodology section of a paper in full.
Where it fits
- Find the latest research papers on a topic like 'diffusion models' or 'quantum computing' directly inside Claude or your AI coding assistant.
- Ask your AI to summarize a specific arXiv paper by its ID, including methodology and results, without copying and pasting anything.
- Monitor a research category like cs.LG (machine learning) for recent papers and get a quick digest of what's new.
- Have an AI read and explain the full text of a paper's PDF, not just the abstract, to help you decide if it's worth reading in depth.