Agent API and MCP

curl "https://arcxiv.org/api/agent/search?q=sparse%20autoencoders&category=cs.LG&sort=relevance&pageSize=5"

curl "https://arcxiv.org/api/agent/papers/2401.12345"

curl -H "Accept: text/markdown" "https://arcxiv.org/abs/2401.12345/markdown"

GET /api/agent/search accepts ordinary research queries, paper titles, authors, arXiv identifiers, and arXiv URLs. A category without a query returns papers from that category.

Results contain full abstracts, normalized metadata, citation counts when OpenAlex has matched the record, assessment signals when available, and stable ArcXiv/arXiv links. Citation counts are refreshed daily for recent papers and are never treated as public quality.

GET /api/agent/papers/{arxivId} returns title, abstract, authors, categories, dates, DOI, journal reference, version history, assessment signals, citation count, related papers, and canonical links. Both modern and legacy arXiv IDs are accepted; URL-encode legacy IDs containing a slash.

GET /abs/{arxivId}/markdown always returns metadata, abstract, and canonical source links. When the paper declares a Creative Commons license and arXiv provides an HTML rendering, ArcXiv adds best-effort full text with headings, LaTeX math, lists, figures, captions, tables, and quotes.

Conversion runs through ArcXiv's private rate-limited arXiv gateway and is cached by exact paper version. Complex layouts may lose formatting; use the returned arXiv link as the source of record.

A score is nullable when the available source does not support a judgment. Preliminary profiles use metadata and abstract; full-paper profiles use transiently processed paper text. Signals are independent assessments, not peer review, citation counts, or personal relevance.

Scores use a 1–10 scale in half-point increments. Category percentile is calculated independently from the public quality formula. Read the complete research-signals methodology.

The downloadable Node bundle exposes search_papers, get_paper, and get_paper_markdown over MCP stdio. Node.js 20 or newer is required.

curl -o arcxiv-mcp.cjs https://arcxiv.org/downloads/arcxiv-mcp.cjs

{
  "mcpServers": {
    "arcxiv": {
      "command": "node",
      "args": ["/absolute/path/to/arcxiv-mcp.cjs"]
    }
  }
}

Set ARCXIV_API_BASE to use another ArcXiv deployment. The server performs read-only HTTP requests and stores no credentials.

• Search responses are cached at CloudFront by the complete query string.
• Paper, markdown, and discovery responses contain no viewer cookies.
• Public agent reads are limited to 300 requests per five minutes per IP.
• Use pagination and reuse cached responses instead of parallel bulk scraping.
• Contact c@coder.company before sustained bulk use.