Skills and MCP configuration for using arx to inspect, fetch, and search locally cached arXiv papers.

GitHub stars
7

Install

/plugin install arx@arx

Run it inside Claude Code after adding the marketplace with /plugin marketplace add eric-tramel/arx.

Contents

Discovered from source

Skills 3

arx-full-text-search

Use this skill when a user asks to find claims, methods, citations, definitions, equations, related work, or other text inside papers already cached by arx.

Workflow

  1. Use lookup_arxiv_papers first when the user names a specific paper and you do not know whether it is cached.
  2. Call full_text_search with concise keyword-style queries. It searches cached local material only and never contacts arXiv.
  3. Use arxiv_id when the user wants results from one paper. Omit it when the user wants to search across the local arx index.
  4. Choose the narrowest useful scope:
    • default or omitted searches title, metadata, and body text, excluding bibliography.
    • titles searches paper titles only.
    • bibliography searches citation records and .bib or .bbl material only.
    • all searches everything.
  5. Read the returned snippets, scores, source paths, and line ranges. Use them as grounded evidence and cite the paper ids or file locations in your answer when helpful.

Notes

  • Empty results may include a note that explains whether the paper is uncached, metadata-only, or missing bibliography/body material.
  • If the paper is missing source or body text, use arx-paper-fetch to queue the needed material before retrying.
  • Prefer a few targeted searches over one broad natural-language question.

arx-paper-fetch

Use this skill when a user wants a paper added to the local arx cache, needs PDF or TeX source material, or needs to know how arx downloading works.

Workflow

  1. Start with lookup_arxiv_papers for the target ids. Confirm which material is already cached and which material is missing.
  2. Call fetch_arxiv_paper only for missing material that is actually needed. It queues work through arxd and returns immediately with a job id.
  3. Tell the user that queued work runs in the background. Do not assume the PDF or source is ready until the queue says so.
  4. Call get_arxiv_download_queue_status with the returned job_id to inspect one job, or with include_finished: true to see completed and failed jobs after arxd exits.
  5. After a completed download, use lookup_arxiv_papers again to refresh readiness, then use full_text_search when the task requires paper text.

Notes

  • arxd enforces the shared arXiv request delay across CLI and MCP clients.
  • Finished job records are durable across arxd restarts, so include_finished: true is the reliable default for follow-up checks.
  • Do not queue downloads when metadata is enough for the user's task.
  • If a job fails, report the failure reason from the queue status and avoid retry loops unless the user asks.

arx-paper-metadata

Use this skill when a user asks about one or more arXiv papers and you need the paper identity, abstract, metadata, local cache state, or the best next arx action.

Workflow

  1. Normalize the user-provided arXiv ids before calling tools. Preserve version suffixes only when the user explicitly asks for a specific version.
  2. Call lookup_arxiv_papers first. It returns metadata, abstract text, local material readiness, cache paths, and per-paper errors.
  3. Treat the lookup result as the source of truth for whether metadata, PDF, source, and index material are already present locally.
  4. If the user asked for a no-network check, pass fetch_missing_metadata: false. Otherwise, lookup_arxiv_papers may fetch missing metadata but never downloads PDFs or source archives.
  5. Use the returned next_tool or readiness fields to decide whether to search cached text or queue a download.

Notes

  • Batch related paper ids in one lookup_arxiv_papers call.
  • Cached metadata should still be used when a network refresh fails.
  • Do not call fetch_arxiv_paper just to inspect metadata.
  • Report per-paper lookup errors separately instead of treating a partial batch failure as a total failure.

MCP servers 1

Sponsors

Details

Marketplace
eric-tramel/arx
Source
plugins/arx
Author
Eric Tramel
Category
Developer Tools
License
MIT