I run a real business out of a markdown AI OS. Three lessons after months of living in it.

Most AI OS builds I see start with automations. Mine started with a knowledge base, and that order turned out to be the whole game.

What I actually run: a plain markdown vault, no database, connected by links, with a team of named specialist agents. An orchestrator that routes, a researcher, a journal writer, a data analyst, and a few more. I drive it from Claude Code, Codex and Telegram, and it holds the real context for operations at my company.

Three things I would tell anyone building one:

1. Build the brain before the hands. A capture habit that turns every input into linked notes beats a pile of clever automations that have no context to act on.

2. One fact, one file. The day the same fact lives in three places, your agent gives you three different answers. Link, do not copy.

3. Stay host-portable. I can open the same vault in Claude Code, Codex or a chat-only LLM. No lock-in means no single tool can ever strand my system.

The trap I see most: handing agents real keys before the context is good enough to trust them. Build the knowledge first, hand over autonomy slowly.

Curious what everyone else runs. What is the one piece of your AI OS you would not give up?

2 comments