# KanBots

> A local-first kanban board that dispatches parallel coding agent CLIs on each card, each in its own git worktree.

KanBots is an MIT-licensed desktop app that turns a project folder into a kanban board where each card can be handed to a coding agent CLI such as Claude Code, Codex, Gemini, Cursor, Copilot or Amp. Each run is isolated in its own git worktree. The latest GitHub release listed is v1.2.0, and a separate hosted Cloud offering exists for teams.

## What It Is

KanBots is a kanban-style orchestration layer for AI coding agents. You pick a folder containing a git repository, and it creates a `.kanbots/` directory holding a SQLite database, config and worktrees. Issues can be local, stored in SQLite, or GitHub issues driven through Octokit. Each card can be dispatched to an agent, with the live thread streaming tool calls and results into the UI.

## How an Agent Run Works

Dispatching a card creates a per-run worktree branched from the default branch and spawns the chosen agent CLI against it. If the agent needs a decision, the run pauses and presents numbered options; replies accept slash commands such as /spec, /review and /split. When a run finishes, the worktree can be previewed through its dev server, promoted to a real commit, opened as a draft PR in GitHub mode, or discarded. A pre-push hook stops agents from pushing on their own.

## Autopilot and Personas

Autopilot has two modes. Feature-dev cycles through a roster of personas, with up to four parallel slots, and agents split parent issues into subtasks. QA runs configurable checks (typecheck, tests, lint, build, e2e) and dispatches fix runs for failures. Sessions stop on completion, manual stop, or a session cost budget. Personas are named system-prompt lenses; built-ins ship with the app and custom ones are stored locally.

## Agents, Integrations and MCP

The project lists 11 supported agent CLIs, including any ACP-compatible CLI, each reusing its own existing authentication. Sentry error groups can be imported onto the board, and a bundled `kanbots-mcp-server` exposes the board over Model Context Protocol so MCP-aware clients like Cursor or Claude Desktop can drive it. Cost analytics track per-run and per-card spend with configurable caps.

## Setup Path

It can be launched with `npx kanbots`, installed from packaged macOS, Windows and Linux builds, or run from source with Node 20+ and pnpm 10+. At least one supported agent CLI must be on the PATH. The README notes macOS and Windows builds are currently unsigned.

## Features
- Parallel agent dispatch, each in its own git worktree
- Autopilot feature-dev and QA modes
- Decision prompts with numbered options and slash commands
- Built-in and custom personas
- Live cost analytics with per-run and session caps
- Local-first SQLite storage with no telemetry
- GitHub Issues mode with draft PR promotion
- Branch preview via worktree dev server
- Sentry issue import
- Bundled MCP server
- Pre-push hook preventing agent pushes
- Support for 11 agent CLIs including ACP-compatible ones

## Integrations
Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot CLI, Amp, OpenCode, Droid, CCR, Qwen Code, GitHub, Sentry, Model Context Protocol, Claude Desktop, Slack

## Platforms
WINDOWS, MACOS, LINUX, IOS, WEB, API, VSC_EXTENSION, CLI

## Pricing
Open Source, Free tier available

## Version
1.2.0

## Links
- Website: https://kanbots.dev
- Documentation: https://www.kanbots.dev/docs/getting-started
- Repository: https://github.com/leodavinci1/kanbots
- EveryDev.ai: https://www.everydev.ai/tools/kanbots