Arxiv2md

arxiv2md

Convert arXiv papers to LLM-ready Markdown through the online service at https://arxiv2md.org. The service parses arXiv's native HTML, not PDFs. Do not install or use the local arxiv2md CLI for this skill.

Standard Workflow

1. Create an output directory for the arXiv ID.
2. Call the online Markdown endpoint with:

• frontmatter=true
• remove_refs=false
• remove_citations=false
• remove_toc=true

1. Save the raw API result as .md.
2. Run scripts/clean_arxiv2md_markdown.py to produce .cleaned.md.
3. Return both paths and note remaining conversion artifacts.

Prefer keeping citations. remove_citations=true can leave awkward empty text such as Shor , Steane , and Knill.

REST API

Base URL: https://arxiv2md.org

No auth required. Rate limit: 30 requests/minute.

Recommended Markdown Request

curl "https://arxiv2md.org/api/markdown?url=2501.11120&frontmatter=true&remove_refs=false&remove_citations=false&remove_toc=true"

Returns raw Markdown as plain text.

JSON Metadata

curl "https://arxiv2md.org/api/json?url=2501.11120"

Returns { "arxiv_id", "title", "source_url", "content" }.

Cleanup Script

Use the bundled cleanup script after saving API output:

uv run python \scripts\clean_arxiv2md_markdown.py .\1708.02246.md -o .\1708.02246.cleaned.md

When arxiv2md frontmatter mixes affiliations into the author list, pass corrected values:

uv run python \scripts\clean_arxiv2md_markdown.py .\1708.02246.md -o .\1708.02246.cleaned.md --author "Christopher Chamberland" --author "Michael E. Beverland" --affiliation "Institute for Quantum Computing and Department of Physics and Astronomy, University of Waterloo" --affiliation "Station Q Quantum Architectures and Computation Group, Microsoft Research"

The cleanup pass handles common arxiv2md/ar5iv artifacts:

• Removes the generated TOC, where math alt-text is often noisy.
• Fixes duplicated Abstract text.
• Corrects common ar5iv visual math alt-text around simple inline formulas, including duplicated variables such as v 𝑣 v, hyphenated variables such as distance- d 𝑑 d, and one-variable numeric expressions such as d + 1 𝑑 1 d+1 or d = 5 𝑑 5 d=5.
• Normalizes ar5iv delimiter fragments that many KaTeX renderers reject, such as \big{(}, \big{)}, \big{\{}, and \big{|}.
• Converts Figure: ... plus Refer to caption: URL blocks into Markdown images.
• Repairs double-wrapped display math such as $$ $\displaystyle ...$ (1) $$.
• Tidies theorem-like headings such as Definition 1 ..
• Converts definition/proof markers from Markdown headings into bold paragraphs, and removes empty duplicate definition anchors emitted by ar5iv.

Parameters

Use these query parameters for both endpoints unless noted:

| Param | Recommended | Description | | --- | --- | --- | | url | required | arXiv URL or ID, such as 2501.11120v1 or https://arxiv.org/abs/2501.11120 | | remove_refs | false | Keep bibliography/references section | | remove_toc | true | Remove generated table of contents | | remove_citations | false | Keep inline citation text | | frontmatter | true | Prepend YAML metadata, /api/markdown only |

Notes

• The online service may still emit ar5iv artifacts in figure captions, footnotes, or complex math. Preserve the raw file beside the cleaned file for comparison.
• The cleanup script intentionally handles only conservative formula artifacts. For complex formulas, inspect the cleaned Markdown against the raw file or arXiv HTML before making semantic corrections.
• If a paper has broken or missing arXiv HTML, report the API failure and do not fall back to pandoc unless the user asks.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DeviesVi
• Source: DeviesVi/skills
• License: MIT

Install and usage instructions live in the source repository linked above.