Search arXiv papers, get metadata/abstracts, and list categories via mcp-simple-arxiv (uvx/Python)

mcp-simple-arxiv v0.6.0 — arXiv paper search, metadata, and full-text via MCP (Python/uvx)

Package: mcp-simple-arxiv on PyPI (by Andy Brandt) Transport: stdio (via FastMCP 2.14.7) Install: uv venv /tmp/arxiv-env && uv pip install --python /tmp/arxiv-env/bin/python mcp-simple-arxiv Run: /tmp/arxiv-env/bin/mcp-simple-arxiv Dependencies: fastmcp, feedparser, httpx, beautifulsoup4, docling (heavy — pulls in PyTorch + RapidOCR for PDF→text)

5 tools — search, retrieve, and convert arXiv papers

Critical gotchas

1. `sort_by` values are SNAKE_CASE — use "submitted_date", "updated_date", "relevance" (not "submittedDate"). Helpful error message on wrong value.
2. `get_full_paper_text` is EXTREMELY SLOW — downloads PDF, then uses docling + RapidOCR (PyTorch CPU) to convert. First call downloads ~40MB of OCR models. Timed out at 60s in my test on "Attention Is All You Need". Not practical for quick agent workflows — use get_paper_data for abstracts instead.
3. `cat:` prefix in query — category filtering works via arXiv's query syntax. Example: "cat:cs.AI large language model" searches only in the cs.AI category.
4. `date_from` filter works — ISO date string like "2026-01-01" narrows results to papers after that date.
5. Nonexistent paper IDs — returns graceful error text "Error calling tool 'get_paper_data': Paper not found: 9999.99999" (not MCP crash).
6. Default search returns by relevance — the order matches arXiv's relevance scoring. Use sort_by: "submitted_date" with sort_order: "descending" for newest-first.
7. Search returns total count — "Found 469984 total results, showing first 3" — useful for understanding result set size.
8. Paper IDs use arXiv format — e.g. "1706.03762" (Attention Is All You Need), "2303.08774" (GPT-4 Technical Report).
9. Shutdown crash is cosmetic — after clean disconnect, the server throws TypeError: An asyncio.Future, a coroutine or an awaitable is required due to Python 3.14 asyncio compatibility. Does NOT affect tool execution.
10. Heavy install — docling pulls PyTorch, RapidOCR, and many ML dependencies. Total install is ~500MB+ but only get_full_paper_text needs them; search and metadata work fine without.

Performance

• list_categories: 2ms (local, cached)
• search_papers: 604-3048ms (network-bound, varies by query complexity and arXiv API load)
• get_paper_data: 638-3032ms (network-bound)
• get_full_paper_text: >60s TIMEOUT (PDF download + OCR conversion)
• search_papers with sort_by: "submitted_date": 604ms (faster than relevance-ranked search)

Verified test trace (13 calls, 11 OK + 1 timeout + 1 correct error rejection)

OK  list_categories                           2ms     4430ch  — full arXiv taxonomy (cs, math, physics, etc.)
OK  search("transformer attention", n=3)      656ms   1433ch  — 469984 results, top 3 with previews
OK  search("cat:cs.AI LLM agent", n=3)        3048ms  1493ch  — category-filtered search
OK  search("MCP model context protocol", n=3) 2879ms  1493ch  — keyword search
OK  search(sort_by="submitted_date", n=2)      604ms  1088ch  — newest-first sorting works
OK  search(date_from="2026-01-01", n=3)        2925ms 1492ch  — date-filtered to 2026 only
OK  search("xyznonexistent", n=2)              2970ms 36ch    — "No papers found matching your query."
OK  search(sort_