Skip to content

AlphaXiv: Research Papers Without the Token Waste

2 min read
Claude Code AI Tools Research

Today I realized I’ve been wasting tokens every time I upload a research paper to Claude.

I read a lot of papers. The workflow was always the same: find a paper on arXiv, download the PDF, upload it to Claude, wait while it chews through 20+ pages of dense LaTeX-rendered content. That’s easily 30,000-50,000 tokens per paper just for ingestion. And most of the time, I only need the core contribution: what problem, what approach, what results.

Then I found the AlphaXiv paper lookup skill for Claude Code. It pulls pre-built, machine-readable summaries from any arXiv paper. No PDF upload. No token burn.

How It Works Under the Hood

The tech is surprisingly simple. AlphaXiv maintains two public endpoints that serve paper content as plain markdown:

# Get a structured overview (problem, approach, results)
curl -s "https://alphaxiv.org/overview/{PAPER_ID}.md"

# Get the full paper text as markdown
curl -s "https://alphaxiv.org/abs/{PAPER_ID}.md"

No authentication. No API key. No rate limiting that I’ve hit. The overview endpoint returns a distilled, structured analysis optimized for LLM consumption. The full text endpoint gives you the complete paper content converted from PDF to markdown.

Why this matters for token cost: A raw PDF upload forces Claude to process the entire document layout, figures, references, and formatting artifacts. The markdown version strips all of that. You get clean, structured text that Claude can parse in a fraction of the tokens.

Setting It Up

Install the skill in Claude Code:

claude install-skill https://www.alphaxiv.org/skills/alphaxiv-paper-lookup/SKILL.md

Then use it with any arXiv paper. All of these formats work:

# Full URL
Explain https://arxiv.org/abs/2401.12345 with /alphaxiv-paper-lookup

# Just the paper ID
Explain 2401.12345 with /alphaxiv-paper-lookup

# AlphaXiv URL
Explain https://alphaxiv.org/overview/2401.12345 with /alphaxiv-paper-lookup

The skill extracts the paper ID regardless of input format, fetches the markdown overview, and gives you back the problem statement, approach, and key results in a structured format.

My Workflow Now

Before: download PDF, upload to Claude, burn 40k+ tokens, wait, get a summary.

After: paste the arXiv link, get the same summary in seconds with minimal token cost.

For deeper dives, the skill falls back to the full text endpoint. If neither is available (rare for popular papers), it tells you to grab the raw PDF as a last resort. But in practice, most papers I care about already have the markdown version ready.

What I Learned

  • AlphaXiv serves any arXiv paper as clean markdown via two public, unauthenticated endpoints
  • The overview endpoint gives you a structured summary (problem, approach, results) optimized for LLMs
  • Raw PDF uploads waste 30,000-50,000 tokens on layout, figures, and formatting that adds no semantic value
  • The skill auto-extracts paper IDs from arXiv URLs, AlphaXiv URLs, or bare IDs
  • 100% free, no API key, works with Claude Code out of the box

Reading a lot of research papers with Claude? I’d love to hear how you’re managing the workflow. Reach out on LinkedIn.

Back to top