# Kanbots
Kanbots is a local-first Kanban board UI that turns every task card into a dispatch slot for an AI coding agent CLI. Drop a folder, get a board, then send each card to Claude Code, Codex, Gemini, Cursor, Copilot, Amp, OpenCode, Droid, CCR, Qwen, or any ACP-compatible CLI — each in its own git worktree, running in parallel.
## How it works
- Five-column Kanban (Backlog → Done) plus an Inbox for unlabeled cards.
- Local-first issues stored in SQLite by default. Optional GitHub mode mirrors moves as `status:*` label edits and can open draft PRs.
- 11 agent CLIs supported, each authenticated via its own existing CLI auth (Kanbots calls the binary already on your `PATH` — no separate sign-in).
- Each agent run is isolated in a per-run worktree; a pre-push hook prevents agents from pushing.
- Live agent thread: every `tool_use` / `tool_result` streams into the UI. Decision prompts pop up; click an option, run continues.
- "Promote" lands an agent's worktree as a real commit or draft PR.
- Branch preview spins up the worktree's dev server in one click.
- Sentry import pulls error groups onto the board for one-click triage by an agent.
- Ships an `kanbots-mcp-server` so Cursor, Claude Desktop, or any MCP client can drive the board.
- "Autopilot" lets the system split tasks across parallel slots and self-check while you sleep.
## Why it matters
This is the **multi-agent orchestration UI** that the AI coding tool ecosystem has been missing. Single-agent workflows (one Claude Code session) hit a wall when work is parallelizable. Kanbots solves the orchestration layer:
- Visual surface for what each agent is doing (vs N terminal windows).
- Worktree isolation prevents agents from stepping on each other.
- CLI-agnostic — you keep your auth, your model choice, your billing per provider.
- MCP exposure means other AI tools can drive the board, not just humans.
For a knowledge worker running a fleet of specialized agents, this is the canonical pattern for parallel task dispatch in code domains. The same pattern (board + isolated execution + promote step) maps onto content pipelines, research pipelines, and any place where a job graph can fan out.
## Implications
- Validates the **fan-out + single-promote** safety pattern for multi-agent work (read-only parallel work, single trusted write at the end).
- Validates **MCP as the universal driver protocol** — Kanbots is both an MCP server (board-as-tool) and an MCP-client consumer of agents.
- Confirms the trend: **the IDE is becoming a Kanban board over a swarm**, not a single editor with a chat sidebar.
- Reusable lens for OSK: parallel-skill execution (e.g. content pipeline, vault maintenance batch) could borrow the worktree-isolation + promote pattern.
## References
- Website: https://www.kanbots.dev/
- Docs: https://www.kanbots.dev/docs/getting-started
- Learn: https://www.kanbots.dev/learn
- GitHub: https://github.com/leodavinci1/kanbots
## Related
-