Show HN: Scry, programmable internet search w/ congestion pricing

一句话概要

Programmatic internet research. Run programs over billions of internet documents, for fun and profit.

原文开头节选

People search for pages; agents can query the underlying records. The question an agent actually has is rarely ten links. It might need every page that mentions this compound but not that patent, grouped by domain, since March . Scry runs that as a query instead of assembling it from repeated searches and scraped results. Each statement is read-only and runs under a deadline, a memory ceiling and a row cap on what returns; what a relation can answer quickly is on its schema entry.

Only the MCP server: https://mcp.scry.io , in ChatGPT, Claude, Claude Code, Codex, Cursor, or any MCP client. The sign-in creates your account. Nothing runs on your machine.

On Google’s DeepSearchQA, 900 research questions graded exact-set, an agent doing its research through Scry answers 71.8% fully correct on a corpus that is still under half of Common Crawl; Google’s Kaggle leaderboard reports its Gemini Deep Research Agent at 66.1%. Same judge model, same grading prompt, and every question, grade and tool-call transcript is on the DeepSearchQA page .

(以上为原文节选,完整内容见下方”原文来源”)

这条动态今日登上 Hacker News 首页(37 赞 / 16 评论,来源 scry.io)。技术雷达每日自动聚合 AI 工程、后端架构、DevOps 方向的前沿动态;相关工程落地可浏览下方的相关服务与延伸阅读,或直接与我们团队交流。