good-eye

Problem

AI coding tools build fast but design generically. There was no way to give an AI agent a curated sense of taste: the specific sites, palettes, and layouts worth borrowing from, available exactly when it is designing something.

Approach

Built good-eye: a local database of full-page screenshots, component crops, palettes, typography, and a written review for every site or app worth learning from. It exposes an MCP server so Claude (Code and Desktop) and Codex can search the library directly, plus a skill and hook that pull relevant references in automatically whenever you ask an agent to design something. Source of truth is plain files; a derived SQLite index powers search.

Outcome

An MCP-native design memory: ask an agent to design a page and it consults a real, curated library instead of guessing. Files-are-truth, self-hosted.

Stack

TypeScript, MCP, SQLite/FTS5, Claude Code skills and hooks

Team

Solo