arXivMCP Server

by Joseph Blazick

Added to Onei Jul 29, 2026

v0.6.2

Repository Docs

Web & Search

Summary

Search arXiv, download papers, and read their full text from an AI client — research without the PDF shuffling.

Features

Search arXiv by query, category, and date range
Download and read full paper text, not just abstracts
Local paper library for repeated reference

Installation

Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.

Any MCP-compatible agent

Most agents (Claude, Cursor, Windsurf, VS Code, and more) read a standard mcpServers configuration.

Open your agent's MCP configuration file.
Merge the snippet below into it.
Restart the agent — the "arXiv" tools will be available.

{
  "mcpServers": {
    "arxiv": {
      "command": "uvx",
      "args": [
        "arxiv-mcp-server"
      ]
    }
  }
}

Claude Code

Add the server with a single CLI command.

claude mcp add arxiv -- uvx arxiv-mcp-server

Claude Desktop

Add the server to claude_desktop_config.json (Settings → Developer → Edit Config).

{
  "mcpServers": {
    "arxiv": {
      "command": "uvx",
      "args": [
        "arxiv-mcp-server"
      ]
    }
  }
}

Cursor

{
  "command": "uvx",
  "args": [
    "arxiv-mcp-server"
  ]
}

Windsurf

{
  "command": "uvx",
  "args": [
    "arxiv-mcp-server"
  ]
}

VS Code

{
  "command": "uvx",
  "args": [
    "arxiv-mcp-server"
  ]
}

Description

The arXiv MCP server gives an assistant direct access to the preprint archive that most machine learning, physics, and mathematics research passes through first.

What it does

Search by query, author, category, and date range, returning structured results
Download papers and convert them to readable text, so the assistant works from the full paper rather than the abstract
Local library. Downloaded papers are stored, so re-reading one across sessions does not re-fetch it.
Reading tools for pulling out specific sections — methodology, results, related work

Why full text matters

Abstracts are marketing. The interesting details — what the experiment actually measured, what the limitations section admits, what the ablation shows — are in the body. An assistant with the full text can answer questions the abstract cannot.

Realistic uses

Keeping up with a fast-moving area, checking whether an idea has already been published, extracting the specific hyperparameters a paper used, and comparing several papers' approaches to the same problem.

Installed with uvx arxiv-mcp-server. No API key required — arXiv's API is open, though it asks that you be reasonable about request rates.

Related MCP Servers

MCP: Altmetric MCP Server

by Altmetric

Altmetric's official MCP server for research attention data — news, policy documents, patents and social mentions of a paper — over a hosted OAuth endpoint or locally with your own API keys.

Web & Search

MCP: Kagi MCP Server

by Kagi

Official Kagi MCP server: ad-free web, news, video and image search plus page extraction as markdown, via hosted or local setup.

Web & Search

MCP: Scry

by Scry

Remote MCP server that lets an agent run read-only SQL over roughly 164 billion indexed documents from Reddit, Hacker News, arXiv, Wikipedia, Stack Exchange and prediction markets.

Web & Search

MCP: Patent & Literature Search

by Patsnap

Patsnap's MCP server for searching global patents across WIPO, EPO, USPTO, CNIPA, JPO and KIPO alongside scientific literature, returning records as AI-readable Markdown.

Web & Search

Browse all MCP servers →