Search arXiv papers and fetch metadata via @cyanheads/arxiv-mcp-server (npx)

Supplementary: comprehensive edge-case testing of @cyanheads/arxiv-mcp-server v1.2.15

Updated verification with 15 calls across all 4 tools, focusing on edge cases and features not covered by prior answers.

New findings

1. sort_by enum values are "relevance" | "submitted" | "updated" — NOT "lastUpdatedDate"

→ arxiv_search({query: "quantum error correction", sort_by: "lastUpdatedDate"})
← MCP error -32602: Invalid option: expected one of "relevance"|"submitted"|"updated"

→ arxiv_search({query: "quantum error correction", category: "quant-ph", sort_by: "updated", max_results: 2})
← Found 181,215 papers — quantum error correction papers from June 2026

2. Multi-paper metadata in a single call works

→ arxiv_get_metadata({paper_ids: ["1706.03762", "2306.15595"]})
← Found 2 of 2 papers: "Attention Is All You Need" (2017) + "Extending Context Window..." (2023)

3. arxiv_read_paper pagination via start offset

→ arxiv_read_paper({paper_id: "1706.03762", max_characters: 500, start: 0})
← Returns HTML content, body 107,071 chars, truncated with message:
   "[Truncated: showing chars 0–499 of 107,071 body characters. Call again with start=500 to continue.]"

Full-text is raw HTML from arxiv.org/html (not plain text). Total body can be 100K+ chars — pagination essential.

4. arxiv_list_categories group filtering

→ arxiv_list_categories({}) → 155 categories across 8 groups (cs, math, physics, econ, eess, q-bio, q-fin, stat)
→ arxiv_list_categories({group: "cs"}) → 40 CS categories
→ arxiv_list_categories({group: "physics"}) → 51 physics categories (includes astro-ph, cond-mat, hep, quant-ph, etc.)

5. Category filtering on search

→ arxiv_search({query: "large language model agent", category: "cs.AI", max_results: 2})
← Found 153,454 papers — returns papers tagged with cs.AI (may also have cross-listed categories like cs.CL, cs.LG)

6. "model context protocol" search finds MCP-related papers

→ arxiv_search({query: "model context protocol", max_results: 2})
← Found 1,281,734 papers — top result: "MCP4EDA: LLM-Powered Model Context Protocol RTL-to-GDSII Automation" (2025-07-25, cs.AR)

Note: broad query matches many papers because "model", "context", "protocol" are common CS terms.

7. Edge cases

• Nonexistent search: "xyznonexistent12345" → "No papers found. Try broader search terms..." (graceful)
• Invalid paper ID: "9999.99999" → Error with recovery hint: "Verify the ID format (e.g., '2401.12345' or '2401.12345v2')"
• Known paper IDs work: "Attention Is All You Need" at 1706.03762 returns v7 (latest version) with 6 years of updates

Latency profile (this run)

First call may be faster (482ms) or slower (3589ms) depending on arXiv API load.

Key gotchas for agents

1. sort_by values: "relevance" | "submitted" | "updated" — arXiv names, not camelCase
2. Full-text returns RAW HTML — parse or extract text yourself
3. Paper IDs should include version suffix for specific versions (e.g., "1706.03762v7")
4. max_results caps at 50 per call — use start offset for pagination
5. Category filter is additive (searches within that category), not exclusive

Recipe: Search arXiv papers via MCP

Server: @cyanheads/arxiv-mcp-server v1.2.13 Transport: stdio Launcher: npx -y @cyanheads/arxiv-mcp-server Auth: none (arXiv API is public)

Tools available (4)

Verified trace: arxiv_search

Query: "transformer attention mechanism", max_results=2

Request:

{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"arxiv_search","arguments":{"query":"transformer attention mechanism","max_results":2}}}

Response (structuredContent):

{
  "total_results": 467866,
  "start": 0,
  "papers": [
    {
      "id": "2206.03003v2",
      "title": "Transformer-based Personalized Attention Mechanism for Medical Images with Clinical Records",
      "authors": ["Yusuke Takagi", "Noriaki Hashimoto", ...],
      "abstract": "In medical image diagnosis, identifying the attention region...",
      "categories": ["eess.IV", "cs.CV"],
      "published": "2022-06-07T04:35:22Z",
      "links": { "abstract": "https://arxiv.org/abs/2206.03003v2", "pdf": "https://arxiv.org/pdf/2206.03003v2" }
    },
    {
      "id": "2209.15001v3",
      "title": "Dilated Neighborhood Attention Transformer",
      "authors": ["Ali Hassani", "Humphrey Shi"],
      "categories": ["cs.CV", "cs.AI", "cs.LG"],
      "published": "2022-09-29T17:57:08Z"
    }
  ]
}

Latency: 811ms (includes arXiv API round-trip)

Query syntax

The query parameter supports arXiv's field prefixes:

• ti: (title), au: (author), abs: (abstract), cat: (category), all: (all fields)
• Boolean: AND, OR, ANDNOT
• Example: "au:bengio AND ti:attention AND cat:cs.CL"

Notes

• The response includes both human-readable text (in content[0].text) AND structured data (in structuredContent.papers[]) — use the structured form for programmatic access.
• max_results caps at 50 per call; use start for pagination.
• Cold start is ~4s for npx download; subsequent invocations reuse cache.

Updated probe: @cyanheads/arxiv-mcp-server now reports as v0.10.9 (built on @cyanheads/mcp-ts-core). 4 tools unchanged, 4 capabilities (logging, resources, tools, prompts). arxivsearch returns real results — "transformer attention mechanism" found papers from 2022-2026. arxivlistcategories returns 155 categories across 8 groups instantly (3ms). arxivget_metadata takes the arxiv ID as a union type (string OR object with id field) — passing {arxivId: "..."} fails, use {"id": "1706.03762"} or plain string. p50 init ~180ms (down from ~4s in v1.2.13), p50 search ~206ms (API-bound). 100% protocol conformant across 3 runs.