Even the sharpest AI helpers benefit from a reliable overseer.

Hey, over at MakerX, we kept hitting this snag: fantastic AI agents for coding, but nothing to tie them together smoothly in everyday development routines. Stuff ended up fragmented in various windows, with no common memory or proper checkpoints for oversight. That's why we created Kagan – it's like a command-line driven board for managing tasks, steering AI agents through every phase of a job. From outlining ideas to executing, checking, and integrating – everything stays connected without dropping details.

What's key here? It avoids over-automating. Think of Kagan as a versatile toolkit for blending AI into your coding process – you decide how much independence to give each individual assignment, rather than locking into one style for the whole endeavor.

Breaking It Down into Phases

Step One: Outlining the Goal

Start by jotting down your project needs. Kagan organizes this into a clear entry on its task board. Execution holds off until you give the thumbs up. For each item, pick between fully automated mode (let it roll independently) or collaborative mode (stay involved) – and feel free to blend them across the board as needed.

Step Two: Execution Zone

At the heart, a background process launches agents in separate git workspaces – keeping things conflict-free even with multiple jobs running side by side. Through the Agent Communication Protocol (ACP), it handles interactions with 14 different agents, including options like Claude Code, Codex, Gemini CLI, Goose, OpenHands, Amp, and several others.

Step Three: Final Assessment

Once done, entries move to a review spot complete with detailed changes, checklists for meeting goals, and summaries whipped up by AI. Hook it up with the GitHub integration to generate pull requests automatically, trigger build tests, and finalize merges – whether you prefer squashing, rebasing, or standard commits – all controllable from wherever you're working.

Keeping You Involved Every Step

This isn't about handing over total control to AI. You pick from two approaches for any given task, tailoring it to fit:

  • Automated Mode – The agent operates quietly in its dedicated git space. Keep an eye on progress via a real-time feed, jump in with chat adjustments if something shifts, and decide the outcome during review before integration. Ideal for straightforward, well-defined jobs where you want things humming in the background.
  • Collaborative Mode – Take the lead yourself. Dive into a live session using your preferred setup (like tmux, Neovim, VS Code, Cursor, Windsurf, Kiro, or Antigravity), work hand-in-hand with the agent, and let Kagan capture the details for easy review later. Perfect for investigative work, big-picture design, or scenarios demanding your direct input.

If the specs are tight and straightforward, go automated. For stuff that's more open-ended or requires your insight along the way, choose collaborative. Your board can handle a mix – adjust based on the task, not the entire setup.

No Need for the Interface If You Don't Want It

You can operate Kagan through any compatible MCP tool without touching its terminal user interface – things like Claude Code, VS Code, Cursor, and so on.