Parallel Agent Orchestration, Disguised as a Kanban Board

What it does

Kanban launches a local web interface from the root of a git repo and detects your installed CLI agent. Each task card on the board spawns an ephemeral git worktree with its own terminal, letting multiple agents work on the same codebase simultaneously without stepping on each other. You create tasks manually or through a sidebar chat, monitor progress via live hooks that surface the latest tool calls on each card, and review diffs or leave comments before committing or opening pull requests.

The interesting bit

Instead of copying the entire repository for every task, Kanban symlinks gitignored files like node_modules into each ephemeral worktree to avoid redundant installs, then uses runtime hooks to display agent activity across hundreds of cards at a glance. Link cards into dependency chains and enable auto-commit, and the board becomes an autonomous pipeline: one task finishes, commits, and automatically kicks off the next.

Key highlights

• Each task card runs in an isolated git worktree, eliminating merge conflicts between parallel agents.
• Gitignored directories are symlinked rather than duplicated, saving disk space and install time.
• Cards can be linked into dependency chains that auto-start when preceding tasks are completed and moved to trash.
• Built-in git interface for browsing history, switching branches, and pushing changes without leaving the app.
• Includes a checkpointing system for diffs and line-level comments sent back to the agent.

Caveats

• The README explicitly warns this is a research preview using experimental features, including bypassing permissions and runtime hooks for more autonomy.
• Symlinking gitignored files means modifying them inside a worktree could cause issues; the authors note you should avoid Kanban if your workflow requires changing those files.
• It is unclear from the README which specific CLI agents are supported beyond a generic “detects your installed CLI agent.”

Verdict

Developers already using CLI agents like Cline who want to parallelize large tasks across isolated worktrees will find this genuinely useful; anyone uncomfortable with experimental autonomy hooks or who rarely runs more than one agent at a time can probably stick with their IDE.