Best for
- Researchers who want AI-assisted arXiv literature discovery
- Users who need local paper storage and full-text or semantic search
- Research workflows able to review external paper content and manage local storage
I
io.github.blazickjp/arxiv-mcp-server
Community
Search, download, and analyze arXiv papers through MCP.
GitHub source repository ↗
64FMRS · C
This Apache-2.0 MCP server, published from the stated GitHub repository at version 0.6.1, is designed for arXiv paper retrieval and research assistance. It is not maintained by arXiv or the MCP protocol owner. Treat retrieved paper text as untrusted input and review AI-generated conclusions before taking action.
ArXiv MCP Server connects AI assistants to the arXiv research repository. It supports paper search, full-text download and reading, local paper listing, LaTeX source access, semantic search over local papers, citation graphs, and research alerts.
Tools
search_papers
Search arXiv papers with query, date, category, result-count, and sorting filters.
download_paper
Download a paper by arXiv ID, trying HTML first and falling back to PDF when needed; stores it locally.
list_papers
List arXiv IDs for papers downloaded locally.
read_paper
Read the full text of a downloaded paper in Markdown with pagination support.
get_paper_latex
Retrieve author-submitted LaTeX source when available.
list_paper_latex_sections
List the section outline of a paper's LaTeX source.
get_paper_latex_section
Read a LaTeX section by section ID or exact title.
semantic_search
Run semantic similarity search across locally downloaded papers; requires pro dependencies.
citation_graph
Fetch references and citing papers through Semantic Scholar; no local download is required.
watch_topic
Create or update a topic watch.
check_alerts
Check topic watches and return papers published since the previous check.
Setup
Install with uv tool install arxiv-mcp-server and launch over stdio with uvx arxiv-mcp-server. Claude Desktop users can also use the published .mcpb bundle; Python 3.11+ is still required. For PDF fallback support for older papers, install arxiv-mcp-server[pdf]. The server can be configured for Streamable HTTP with TRANSPORT=http, HOST, and PORT.
claude_desktop_config.json
{"mcpServers":{"arxiv-mcp-server":{"command":"uvx","args":["arxiv-mcp-server"]}}}
Fit and risk
Not for
- Users who only want browser-based search without local downloads
- Sensitive workflows that cannot accept paper text entering model context
- Deployments requiring proven official maintenance or enterprise authentication
Required permissions
- Access arXiv to search for and download papers
- Read and write papers in the configured local storage directory
- Access Semantic Scholar when using citation_graph
- Listen on the configured local host and port in HTTP mode
Risks and side effects
- Paper content is untrusted external input and may contain prompt-injection instructions
- Combining this server with filesystem, shell, or browser tools increases tool-chain attack risk
- semantic_search, citation_graph, alerts, and related features are experimental and may behave unexpectedly
- arXiv rate limits can cause search requests to fail temporarily
Troubleshooting
- Use uv tool install rather than npm, pnpm, npx, or uv pip install to expose the executable
- Confirm Python 3.11+ is available and run arxiv-mcp-server --help to verify installation
- read_paper and semantic_search require a prior paper download; semantic_search also requires pro dependencies
- Install arxiv-mcp-server[pdf] for older papers available only as PDFs
- If rate limited by arXiv, wait 60 seconds before retrying
- For reverse-proxy HTTP deployments, check ALLOWED_HOSTS and ALLOWED_ORIGINS and retain suitable authentication and network controls
Use cases
Search arXiv by topic, category, or date
Download and read full papers or LaTeX source
Run semantic search across a local paper collection
Inspect citation relationships and monitor research topics
Use research prompts for paper summaries, comparisons, and literature reviews
Supported clients
Claude DesktopFull support
CodexPartial support