KanBots turns a kanban board into an agent dispatcher

KanBots hit the Hacker News front page today with a metaphor that is almost too obvious in hindsight. You already think about work as cards moving across columns on a board. So what if every card on the board was a coding agent that just did itself? That is the whole pitch, and it is open source, MIT licensed, local first, free forever.

Mechanically: each card dispatches a Claude Code or Codex agent into its own git worktree, and they run in parallel, so the board is genuinely doing several things at once rather than one chat at a time. There is an autopilot mode where personas, a product manager, an engineer, a designer, cycle through the backlog, split the work and evolve it on their own. Agents pause to ask you a clarifying question rather than silently rewriting your code, which is the detail that tells you the builders have actually run agents in anger. It tracks cost per run with budget caps, talks to real GitHub issues, and opens draft PRs straight from the board. Everything stays on your machine in SQLite, no telemetry.

The reason the kanban framing works is that it solves the part everyone hand waves about parallel agents: orchestration and visibility. Running ten agents at once is easy. Knowing what ten agents are doing, and stopping the one going off the rails, is the hard part, and a board you already understand is a surprisingly good control panel for it. It joins a crowded shelf of orchestrate-many-agents tools right now, but the open-source-and-local angle gives it a clear lane. Repo and downloads at kanbots.dev.