IDB OS
My agentic setup: everything on this Mac that lets AI agents work for me, what is installed, how the pieces connect, and what was really used.
01 Copilots
The modes I work in. Each one gathers every skill, agent, connector, service, project and tool it uses, and says where it stands: in use, in development, idle or backlog. Open one to see what it is made of and how it runs.
Not in any copilot
Everything installed that no copilot claims. It stays on its own board; to place it, name it in a copilot.
02 The landscape
Seven layers, from me at the top to the results at the bottom. A box with an arrow opens its boards.
How busy it has been
03 Skills, by what they are for
A skill is a written instruction for one kind of task. Claude loads it when that task comes up. Open a card to see what it is for and how it has been used.
04 Agentic flows
How the pieces work together, step by step. A red edge marks a step where I decide. A step with an arrow opens the board behind it.
05 Boards
Every other piece of the setup, one card each. Open a card to see what it is for, how it has performed and whether it is in use.
06 The stack
What makes it all work: the models that think, the models that generate, and the plumbing underneath. Every count comes from the machine; the lines are written by hand in curated/stack.json.
07 Installed, not in use
08 How it is counted
How the numbers are counted
- Everything is read from this Mac: Claude Code's settings, its session transcripts, the services the Mac keeps running, each project's folder. Nothing is typed in by hand except the one-line descriptions and the flows.
- "In use" means used in the last seven days. "Quiet" means used earlier, then not. "Not used yet" means no session has called it since this Mac was set up.
- A copilot or a project has a life stage, read from two dates over the last two weeks. "In use": something of it ran (a skill loaded, an agent run, a connector called, a service alive, a command run). "In development": nothing ran, but a file changed or a session worked in the folder. "Idle": it ran once, then nothing. "Backlog": nothing of it has ever run here and nothing was touched: drafted, or not rolled out.
- On the Stack screen, a thinking model's calls and share come from the session transcripts; a generating model's runs from the ledger its tool keeps (one line per run, only the model and the day are read); the local models keep no log, so their runs say "not recorded". Plumbing counts the pieces of the setup that run on it.
- Every card carries "last used" (the last day it ran) and "last changed" (the last day its files were written, a project's newest file or commit, a service's launch file). Where no date is kept, the card says "not recorded" rather than guess.
- A skill counts as used when a session loads it. A skill whose one-line description is enough to act on can shape a reply without ever being loaded, so its count can read lower than its effect.
- Shares of usage are a weighted estimate (input, cached input and output tokens, times a factor per model family). They compare one part with another. They are no bill.
What this page does not show
- What any memory note, prompt or reply says. The page holds names, counts and dates only.
- Money spent on generation and on subscriptions. Those are kept in each project's own ledger.
- Work done in the Claude desktop and web apps. Only sessions run on this Mac in the terminal leave a transcript here.
The page is a snapshot. Running python3 build.py --publish in the project reads the machine again and republishes it here, at os.idbuldin.com.