blazickjp/arxiv-mcp-server
A Model Context Protocol server for searching and analyzing arXiv papers
3,025 ★241 forksPythonUpdated 9d ago
What you need to know
MCP server for arXiv — search papers, download them, read bounded full text, retrieve original LaTeX by section, follow citation graphs, and maintain research alerts. Runs locally over stdio by default; papers and indexes stay on your machine.
Install
uvx arxiv-mcp-server (requires uv / uvx) uv tool install arxiv-mcp-server (persistent PATH install) Optional extras: 'arxiv-mcp-server[pdf]' (PDF conversion via pymupdf4llm) and 'arxiv-mcp-server[pro]' (local semantic search embeddings) macOS Claude Desktop: bundled .mcpb from latest GitHub release (Apple Silicon / Intel), requires CPython 3.11.x
Usage
- •claude mcp add --transport stdio --scope user arxiv -- uvx arxiv-mcp-server (or codex mcp add arxiv -- uvx arxiv-mcp-server)
- •search_papers with query/categories/sort_by, then get_abstract by paper_id
- •download_paper converts to local Markdown (HTML first, PDF fallback), then page through cached content with read_paper using next_start
- •get_paper_latex -> list_paper_latex_sections -> get_paper_latex_section for section-level LaTeX reading
- •Streamable HTTP: TRANSPORT=http HOST=127.0.0.1 PORT=8080 uvx arxiv-mcp-server
Key features
- ✓14 tools: search_papers, get_abstract, download_paper, read_paper, get_paper_latex (+ sections), citation_graph (Semantic Scholar), export_citations (BibTeX), watch_topic/check_alerts, semantic_search/reindex ([pro])
- ✓Bounded full-text reading with content_length/next_start pagination to avoid context overflow
- ✓Section-level original LaTeX retrieval from validated, size-limited, cached source archives
- ✓7 built-in MCP prompt workflows (research-discovery, deep-paper-analysis, literature-review, etc.)
- ✓Streamable HTTP transport with DNS-rebinding protection and ALLOWED_HOSTS/ALLOWED_ORIGINS
Best for
Researchers and literature-review workflows where an agent must search, read, and cite arXiv papers with bounded context
Caveats
- ⚠Do NOT install via npm/pnpm/npx — an unrelated npm package uses the same name; use uv/uvx
- ⚠PDF fallback requires the [pdf] extra; semantic search requires [pro] and only operates on already-downloaded papers
- ⚠Paper text and LaTeX are untrusted external content — papers can contain prompt-injection-style instructions; use client approval controls
- ⚠Keep Streamable HTTP private unless authentication is provided upstream
Platforms: macOS · Windows · LinuxClients: Claude Code · OpenAI Codex / ChatGPT desktop · Kiro · VS Code · Claude Desktop · Any MCP client
Documentation ↗Reviewed 2026-08-07
Topics
- Stars
- 3,025★
- Forks
- 241
- Language
- Python
- License
- Apache-2.0
- Created
- 2024-11-29
- Last push
- 2026-07-29