I built this for bk1. Well — I was supposed to be working on bk1, and instead I built this. Classic.
I always have a project called "stuff" — a dumping ground for the little things I need to do. I wanted something similar, but where a swarm of Claude Code agents could pick up tasks and implement them autonomously.
I could have learned an existing framework. Building it was more fun, and the result fits exactly how I work.
How it works
Start the app with npm run dev, set up a project, then spawn some workers with the-stuff-worker. They come online and sit there — you can see them as little chips on the board, with an hourly activity chart above the columns.
Tasks start as drafts; moving a card to To Do is the signal that dispatches it. An agent picks it up, works in its own git worktree, implements the change, opens a branch, and files a PR. Tasks can also declare dependencies, so an agent won't touch one until its prerequisites are archived.
Some tasks are research-only — the agent stays in plan mode, writes no code, and spawns follow-up tasks instead. Handy when I'm not sure what I want yet.
The nicest bit is the skill. The repo ships a Claude Code skill that I symlink into ~/.claude/skills, and from then on any Claude session can CRUD the board in natural language — "create a task to fix the x-axis on the activity chart" just works.
It's pretty cool to see the jobs getting dispatched.
I've tried to keep the token bill sane too. Agents reuse the same Claude session across a handful of tasks, and if a task bounces back to TODO it gets routed to the agent that worked on it before — so the context carries over instead of being rebuilt from scratch. There's more to do here: subtasks and tasks linked by dependencies could go to the same agent for the same reason.
Repo's on GitHub if anyone's curious. It's not polished, it's vibe-coded, but it works: github.com/ewoij/the-stuff