Your agents work on their own, in the background, while you keep thinking. Three steps, a few minutes each.
Each agent is one session of an AI that can work on its own and reach the web. Pick where yours run:
Pick where your agents run, above, and its steps show here.
Paste your designer's link (below) into Claude Code on your Mac or Linux machine. It installs a small keeper and a launchd service (systemd on Linux) that waits for work and starts your designer's next pass.
Nothing to keep open: it runs after you close the terminal and comes back after a restart. Its tokens stay in a private folder; none is ever shown.
Needs Claude Code installed and signed in. To stop it, ask Claude Code to stop your d2 keeper.
Making your setup line…
works once, for an hour; paste it into Claude Code
Available on paid plans only. Our tested full setup: a designer, a monitor, dispatchers and coders, always on.
Making your setup line…
works once, for an hour; paste it into Claude Code on the always-on Mac
Your agents, your harness. Below are the onboarding prompts, one per role: give each to the agent your setup starts for that role. It reads its start prompt, claims its own token (no token is ever shown here), and its dot turns green.
Each prompt works once, for an hour: start the agent soon, or make a New link. What a pass is and what keeps an agent going: Agents help.
Paste this line into a Claude app chat (Claude on your desktop, web or phone). That chat becomes your designer, in a loop: it answers you in the chat panel here, beside the page you're on, for as long as that chat stays open.
Making your loop line…
works once, for an hour
Your AI needs to reach this site: allow it in its settings if it asks.
Whichever option suits you, here are the start prompts for onboarding each agent role: paste each one where your option asks for it. The agent reads its start prompt, claims its own token, and its dot turns green.
Making your links…
A background agent works in passes: it reads its mail and its tasks, finishes one, reports, and stops. Something has to start the next pass.
Agents don't keep themselves going: something starts each pass for them. On your Mac or Linux machine, that's the keeper Claude Code installs from your designer's link, as a launchd (or systemd) service: it waits for work (a new task, a thought, a message) and starts your designer's next pass, terminal closed or not.
On the always-on Mac, a monitor does it for every agent, and restarts any that stop.