// the find
midday-ai/packrun
npm for agents and humans
packrun is a scored npm package registry alternative — think 'which date library should I use' answered with real data instead of a Stack Overflow thread from 2019. It pulls live npm stats, bundle sizes, and maintenance signals into a scoring formula, then exposes all of it through a REST API and an MCP server so AI coding assistants can query it mid-session. Squarely aimed at teams tired of making package decisions by gut feel.
The MCP server integration is the best idea here — asking your IDE's AI 'compare axios vs got vs ky' and getting back scored, structured data is genuinely more useful than either asking Claude cold or reading npm pages manually. The scoring formula is explicit and auditable (weights are documented, not a black box). Typesense-backed sub-50ms search is the right call — you don't want a slow lookup inside an AI tool call. The 50+ pre-built category comparisons mean common decisions (ORM, testing framework, logger) are answered without hitting any external API at query time.
The scoring formula treats bundle size and weekly downloads equally (20% each), which will systematically mis-score server-side packages where bundle size is irrelevant — pino gets penalized for nothing. The 'Quality' bucket bundles TypeScript support, ESM, security, and tree-shaking into one 25% weight with no sub-breakdown visible in the docs, so you can't tell why a package scored 72 vs 81. There's no mechanism for human override or community correction — if the formula is wrong about a package (and it will be), there's no escape hatch. Self-hosting requires Typesense Cloud, Redis, and GitHub token to even start; the dependency list is heavier than the README implies.