MCP server exposing arXiv as a browsable virtual filesystem via OAI-PMH + Atom API
Three tools with clear action verbs and reasonable descriptions (avg 156 chars). All have input schemas with types and descriptions. However, output schemas are undocumented, LLMs cannot predict what fields to expect or chain results. Parameter descriptions lack format constraints (e.g., arXiv ID format '2501.12345' vs 'cs/9901002' is mentioned but not formalized). Error handling is minimal, no recovery guidance for rate limits or invalid queries. Tool composition is sound (each does one thing), but missing pagination details and result limits in schema.
Fetch full metadata for an arXiv paper by ID. Returns an embedded metadata.json block plus resource links to all paper files: abstract, full text, LaTeX source, structured references, and per-section files.
List all arXiv subject categories from the live OAI-PMH taxonomy. Pass group to filter by top-level subject area. Valid groups: cs, math, physics, econ, eess, q-bio, q-fin, stat.
Search arXiv and return papers as resource links for direct reading. Field prefixes: ti:(title) au:(author) cat:(category) abs:(abstract) all:(all fields). Boolean operators: AND OR ANDNOT. Example: "ti:transformer AND cat:cs.AI"
Output schemas undocumented. arxiv_search, arxiv_get_paper, and arxiv_list_categories lack documented return types. LLMs cannot predict field names, types, or structure, forcing them to guess or fail on chaining.
Parameter format constraints missing. arxiv_get_paper accepts 'arXiv paper ID, e.g. "2501.12345" or "cs/9901002"' but no regex pattern or enum. LLMs may pass invalid formats like 'arxiv-2501-12345' or 'paper_2501.12345'.
Error handling lacks recovery guidance. Rate limit errors ('arXiv rate limit, wait 3s between calls') are thrown but no tool provides a retry mechanism or backoff strategy. LLMs receive raw errors with no actionable next step.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2025-06-18+ | v2 |
Pagination not formalized in schema. arxiv_search accepts maxResults (default 10, max 50) and start (default 0), but no total_count or next_cursor returned. LLMs cannot determine if more results exist or how to fetch them.
arxiv_list_categories 'group' parameter optional but no guidance on behavior when omitted. Description says 'optional' but does not explain what is returned (all groups? error? empty?).