All of arXiv, as a database your AI agent can query.
Scry arXiv is an MCP server with seven tools. Connect it to Claude, ChatGPT, Cursor, or any MCP client, and your agent runs read-only SQL over 3.0 million arXiv papers: the metadata of every paper, the full text of 2.7 million of them (43 billion tokens), voyage-4-nano embeddings of that text for semantic search, and the OpenAlex citation graph for walking who cites whom. Nothing else of Scry is in the way.
An example
The most cited machine-learning papers posted to arXiv in 2024:
SELECT arxiv_id, title, cited_by_count FROM academic.catalog WHERE has(categories, 'cs.LG') AND startsWith(arxiv_id, '24') ORDER BY cited_by_count DESC LIMIT 5;
2404.19756 KAN: Kolmogorov-Arnold Networks 508 2401.09417 Vision Mamba: Efficient Visual Representation Learning ... 397 2411.06649 A Novel Combined Data-Driven Approach for Electricity Theft 293 2410.08770 Causal machine learning for predicting treatment outcomes 259 2408.00714 SAM 2: Segment Anything in Images and Videos 240
The query ran in 41 ms over the whole catalog (measured 2026-09-19). Full text is one hasAllTokens(text, [...]) away; semantic search is embed the
paragraph you hope exists, then rank every paper by distance to it; citations are a datalog walk. The guide tool carries a worked statement for each.
Connect
Add https://mcp.scry.io/arxiv as an MCP server in Claude or ChatGPT. In Claude Code:
claude mcp add --transport http scry-arxiv https://mcp.scry.io/arxiv
The sign-in creates your account. Nothing runs on your machine. Steps for each client.
Price
Researchers $0. Patrons $100 a month. Teams from $2,000 a month. Agents $0.05 per second of the time box they declare, charged at admission. Every account starts with $5 of credit and no card. Price list.
The rest of the internet
The same account works at https://mcp.scry.io, where the whole of Scry is:
Hacker News, Reddit, the Twitter archive, books, forums, SEC filings, the crawl. Scry in one minute · Sources · Docs · hi@scry.io