Kanbots: Open-source kanban desktop app dispatches parallel AI coding agents per card

Kanbots is an MIT-licensed Electron desktop app that reframes AI coding assistants around a kanban board instead of a chat window. Each card on the board can dispatch a Claude Code or Codex agent into its own git worktree on a dedicated branch, letting users run many agents in parallel while the board streams tool calls, decision prompts, and cost meters in real time. An autopilot mode cycles through personas — product, engineer, reviewer, tester — at a configurable parallelism up to four, with personas spawning subtasks as work is discovered.

The project is local-first by design: SQLite, configs, and worktrees live in a .kanbots/ directory next to the repo, with no cloud account, no telemetry, and no HTTP server. It integrates with GitHub Issues via a personal access token, imports errors from Sentry, exposes the board over MCP for tools like Cursor and Claude Desktop, and uses a pre-push hook so agents cannot publish code on their own. A separate paid tier adds team features — presence, sync, SSO, audit logs — but agents still execute on local hardware.

The pitch is containment and observability for multi-agent development: reviewable decision points instead of silent tree mutations, per-run and per-session cost caps, and a single adapter that normalizes Claude Code and Codex stream formats. It is a concrete attempt to turn parallel AI agents into a manageable system of record rather than a scattering of terminal sessions.