Xyra Sinclair

Hello! Last year, I posted https://www.lesswrong.com/posts/mB2knzKYcrZkTssYQ/show-lw-alignment-scry-1. I want to inform you that Scry has substantially improved. In all of your agent chats that support MCP, should you signup and add mcp.scry.io, your agent can execute extremely expressive query programs--SQL and much of Datalog--over a substantial fraction of the internet (500 TB uncompressed as of September 7, 2026). I've been indexing everything I could imagine a serious person wanting to have programmatic access to in the intelligence explosion. That includes: * the supermajority of forums, blogs, comments, mailing lists, and company sites in the tech/nerd/wealthy supercluster * 90 million full-text academic papers + OpenAlex and its academic citation graph * the 59 million SEC filings * over 94% of reddit * the biggest prediction markets * billions of video transcripts * for vetted clients, some of the highest signal social archives in the world There's 10B Voyage-4 family 2048-dim f16 embeddings and operators to help productively leverage the compositional properties of embeddings. Information about sources and freshness is available through the MCP--you get raw access to the database. I've done substantial work for agent ergonomics so that agents have productive query recipes and guidance when using Scry. I would suggest throwing your deepest curiosities at it.