Arxiv MCP Server gives an AI agent tools for working through arXiv literature. The working loop is paper ID, then outline, then one section, then citations: an agent reads author-submitted LaTeX one section at a time, exports BibTeX from arXiv metadata, and keeps standing topic watches. Downloaded papers, watches, and the optional semantic index stay on local disk, while search, source retrieval, and citation graphs call the arXiv and Semantic Scholar services.
It runs locally over stdio by default, with a Streamable HTTP transport for deployments where stdio is not practical. Recipes and plugin packages cover Claude Code, Codex, Hermes, VS Code, Kiro, and Claude Desktop, and the Claude Code and Codex plugins also install a bundled arXiv research skill. Reads are bounded to 12,000 characters per call by default and page forward with an offset, so one call cannot flood the client's context window.
The server exposes 19 tools, grouped here by job:
| Group | Tools |
|---|---|
| Search | search_papers, get_abstract |
| Full text | download_paper, list_papers, read_paper, get_paper_outline, read_paper_section, search_paper_text |
| LaTeX source | get_paper_latex, list_paper_latex_sections, get_paper_latex_section |
| Citations | citation_graph, export_citations |
| Topic watches | watch_topic, list_watches, check_alerts, unwatch_topic |
| Semantic search (optional) | semantic_search, reindex |
Seven built-in MCP prompts cover research discovery, deep paper analysis, paper summaries, paper comparison, literature reviews, literature synthesis, and research-question formulation. The server supplies the workflow instructions and runs no model of its own. Paper text is treated as untrusted content that may carry prompt-injection attempts.
Features
- Section-level LaTeX reads: retrieves author-submitted LaTeX, lists its section outline, and returns one bounded section at a time
- Bounded full text: converts papers to Markdown (HTML first, optional PDF fallback) and pages through them in fixed-size chunks
- arXiv search: queries by phrase, field, category, date, and sort order, with compact abstract snippets by default
- Citations: fetches references and citing papers from Semantic Scholar and exports BibTeX from arXiv metadata
- Topic watches: saves standing topic watches locally and checks them for papers published since the last check
- Local semantic search: an optional extra indexes downloaded papers for similarity search
- Research prompts: seven MCP prompt workflows, from literature reviews to research-question formulation
- Client packages: plugins for Claude Code and Codex, a Kiro Power, and a Claude Desktop bundle, each with the research skill where supported
- Transports: stdio by default, plus Streamable HTTP with a health check and DNS-rebinding protection
