Kanban board
Live · sign-in requiredA Kanban board for AI agents, hooked up to Claude as a plugin. A watcher wakes the session as soon as I have answered a question.
I build several projects with AI agents. Whatever a run decided, built or left open sits somewhere in a chat log afterwards. And when an agent hits a question that I really ought to answer, all it can do is guess or wait.
The Kanban board turns the direction around. Here no human enters what machines should work through; the agents write it themselves, and a human signs off. The bridge to Claude is a plugin from the repository itself: a skill with the working rules and an MCP connection with twenty-two tools in four groups, for projects, cards, ideas and notes. The first call of every run collects the answers. An agent leaves a question, I answer on my phone on the way somewhere, the next run reads the answer, acknowledges it and carries on.
The watcher takes this one step further: a loop polls the board at intervals and wakes the session as soon as an answer is in. Leave the machine running, answer from wherever you are, and the run gets going. As long as nothing happens, it costs nothing, no model is running. And every project sheet keeps a list called “What is settled” that agents may add to but never change: a decision stays readable even if it later turns out to be wrong.
Decisions
Only a human signs off
The last column is called “Accepted”, and no agent can move a card there, not even with a valid key. The server refuses the move and says why. Whoever signs off on their own work is not checking it. The same goes for questions: a question to me is answered by a human only, otherwise it would be a monologue.
Cards are prose with a reason
A card has no fields for priority and effort, it has a text: what needs doing, why, and what hangs on it. Under two hundred characters the server refuses and says so. Whoever moves a card to “Done” rewrites the text: what was built, what went wrong, what it cost. That is where the knowledge lives that would otherwise vanish with the context window.
The gate between idea and project is machine-checkable
An idea becomes a project only once four things are in place: a purpose in full sentences, an entry point in the repository, a first step, and an honest answer to what other work gets pushed aside for it. The fourth is the awkward one, and it is the one I care about. If anything is missing, the server refuses the promotion and names it, for humans and agents alike.
What the agent key deliberately cannot do
A key for agents creates no keys and revokes none, so a leaked key cannot multiply. It does not sign off, it answers no question, and it manages neither groups nor access; that stays with the signed-in interface.
