What it is
A second brain that lives on the internet, not on your machine. PARA structure, a knowledge graph with WHY-annotated captures, bidirectional links. Full surface: Chrome extension, dashboard, CLI, MCP server, and autonomous agents (graph hygiene, weekly review, daily digest) running as Durable Objects on Cloudflare’s edge.
The opinion
Local-first is the sacred cow of second brains, and I think it’s wrong for the agent era. My computer sleeps; my agents don’t. The brain has to live online so Claude Code, Claude Desktop, ChatGPT (whatever I’m working in) can reach it anytime, from any machine. That’s the whole design.
The second opinion: every capture records why you saved it. Content alone mis-files things: a bookmark could plausibly belong five places. The stated intent is what lets the LLM connect it correctly, and correct connections are what make retrieval work later.
What happened
Three ground-up rebuilds in four months: an MCP-primitives prototype, then a deliberate step back into a local Claude-Code-native version, then the cloud rebuild. It’s launched, multi-tenant, and free. I pay the bills, and almost nobody uses it yet. It’s my exploration project and my daily tool, improving every week. When it’s good enough, I’ll migrate off D1’s limits to Postgres + pgvector and start inviting people properly.
What I learned
Rebuilding three times wasn’t waste. Each version taught me what the next one had to be. And dogfooding is undefeated: the features that survive are the ones I reach for without thinking.
The full build log and the argument against local-first is in Everyone builds local-first second brains. I’m building the opposite.