DEV Community

Cover image for At MakerX, we built Kagan — a conductor for your AI coding agents
Altynbek Orumbayev
Altynbek Orumbayev

Posted on

At MakerX, we built Kagan — a conductor for your AI coding agents

Your best agents still need a conductor.

At MakerX we kept running into the same problem: great AI agents, but no coherent way to orchestrate them across a real dev workflow. Tasks scattered across terminals, no shared state, no structured review gate. So we built Kagan — a keyboard-first Kanban TUI that orchestrates coding agents across the full task lifecycle. Plan. Run. Review. Merge. No context lost between steps.

And critically: it doesn't try to automate everything. Kagan is a Swiss Army knife for AI-assisted development — you choose the level of autonomy per task, not per project.

Three acts

01 — The Brief
Describe what you want built. Kagan turns it into a structured task on a Kanban board. Nothing runs until you approve. Choose AUTO (hands-off) or PAIR (interactive) per task — you can mix both on the same board.

02 — The Stage
The core daemon spawns agents in isolated git worktrees — zero conflicts between concurrent tasks. The Agent Communication Protocol (ACP) manages sessions across 14 supported agents: Claude Code, Codex, Gemini CLI, Goose, OpenHands, Amp, and more.

03 — The Verdict
Tasks land in REVIEW with full diffs, acceptance criteria checklists, and AI-generated summaries. The GitHub plugin auto-creates PRs, runs CI checks, and merges — squash, rebase, or merge commit — from any interface.

Human in the loop, always

This is not an "AI takes the wheel" tool. Two modes, your call per task:

  • AUTO — agent runs in the background in its own git worktree. You watch the live stream, steer mid-run with follow-up chat, and make the final call at REVIEW before anything merges. For clear, bounded tasks where you want async progress.
  • PAIR — you drive. Open an interactive session in your tool of choice (tmux, Neovim, VS Code, Cursor, Windsurf, Kiro, Antigravity), code alongside the agent, and Kagan tracks state and surfaces it at review time. For exploratory, architectural, or anything that needs your hands on it.

Requirements clear and bounded? AUTO. Exploratory, needs your judgement mid-way? PAIR. The board holds both — switch per task, not per project.

The TUI is optional

Any MCP client can drive Kagan without the TUI — Claude Code, VS Code, Cursor, Gemini CLI, Codex, and more. The board, state, and review gate are all accessible via MCP tools.

Try it

uvx kagan
Enter fullscreen mode Exit fullscreen mode

GitHub: https://github.com/kagan-sh/kagan

Would love to hear how this fits into your workflow — especially if you've found the right balance between AUTO and PAIR on a real project.

Top comments (0)