arxiv-mcp

Turn your AI assistant into a live arXiv lab — search papers, pull clean text, follow citations, and stash what matters for later.

*Repo: github.com/sandraschi/arxiv-mcp*

---

Keeping up with arXiv is noisy. You bounce between tabs, PDFs that don’t copy well, and half-remembered paper IDs. arxiv-mcp gives an agent (or you, via a small dashboard) one consistent pipe: find → read → connect → save. When arXiv’s experimental HTML exists for a paper, you get Markdown-friendly text without fighting a PDF first.

Good fit if you: skim new cs.AI / cs.LG drops, deep-read a few papers a week, or want ingested text in a local corpus for RAG and notes.

---

Prerequisites: Python 3.11+, uv, and for the dashboard Node.js (LTS is fine).

A — Use it inside Cursor (or any MCP client over stdio)

git clone https://github.com/sandraschi/arxiv-mcp.git
cd arxiv-mcp
uv sync
uv run python -m arxiv_mcp --stdio

Add a server command pointing at that line (repo root as cwd). You’re done.

B — Browser dashboard + HTTP MCP

uv sync
uv run python -m arxiv_mcp --serve

In another terminal, from the same repo:

.\start.ps1

Open the URL Vite prints (by default http://127.0.0.1:10771). The UI talks to the API on 10770.

---

• *“What’s new in cs.RO in the last day?”*
• *“Pull the full text for this arXiv id as markdown if HTML exists.”*
• *“Who cites this paper / what does it cite?”*
• *“Save this paper into my local corpus for later search.”*

Exact tool names, ports, and stack live in [docs/TECHNICAL.md](docs/TECHNICAL.md).

---

Copy .env.example to .env if you want custom host/port or a Semantic Scholar API key. Most people can skip this at first