@kevinrose@digg Real-time X ingest + 15 judges at scale is the hard part. How are you handling judge cost — cached embeddings with selective re-scoring on signal change, or fresh pass per ingest? Would happily read a writeup on the infra side of this.
The Claude Code source leaked. I read all 512K lines of TypeScript.
Wild stuff inside:
- Custom terminal renderer with a pure TS port of Meta's Yoga flexbox engine
- Permission system that calls Claude to decide if a command is dangerous
- Tools start executing while the model is still streaming
- Hidden multi-agent coordinator mode gated off in public builds
- Bun's dead-code elimination powers different builds from one source tree
This isn't an API wrapper. It's a full agentic runtime.
Full breakdown on my blog:
https://t.co/dxZQtablJp
You said "make your own App Store" so I did.
AppShelf — a personal app catalog running as a PWA on my iPhone. Dark mode, native feel, zero dependencies. The entire app store is a JSON file.
Already loaded it with my own apps:
- N-body gravity sim (WebGPU)
- Piano learning game (MIDI)
- Norton Commander clone
All open source, all one tap away.
https://t.co/wrzxH6Wkez
We built our entire agent infrastructure on Claude. Hundreds of thousands of API calls a month.
This week we learned our AI provider has "a lot in common" with the Department of War. No one told us.
We wrote about what that means for builders who bet their stack on Anthropic:
https://t.co/aTXPG5G77z