Vibe Kanban: The AI Agent Orchestration Layer Your IDE Is Missing
Hook
The bottleneck in AI-assisted development isn’t getting 10X more out of your coding agent—it’s the 15 seconds you spend switching between your kanban board, terminal, GitHub diff view, and browser preview every time you want to review what your agent just built.
Context
We’re living through a strange inflection point in software development. Coding agents like Claude Code, Codex, Gemini CLI, and others can generate working features in minutes, but the human workflow around them hasn’t caught up. You describe a feature in a kanban board, switch to your terminal to run the agent, switch to your IDE to review diffs, switch to your browser to test, switch back to GitHub to create a PR, then finally update your kanban board. Each context switch burns cognitive load and kills momentum.
Vibe Kanban treats this workflow tax as the core problem. Instead of bolting AI onto existing tools, it reimagines the entire development cycle around the assumption that agents do the coding while humans do the planning and review. The result is a unified interface that combines kanban planning, agent execution, diff review, browser preview, and PR creation—all without leaving a single window. It’s not trying to be a better IDE or a better project management tool. It’s trying to be the orchestration layer that sits above both, purpose-built for the plan-execute-review cycle that dominates AI-assisted development.
Technical Insight
The architecture reveals some clever decisions about how to isolate agent work without creating merge conflicts. Vibe Kanban is a dual-process application: a Rust backend handles git operations, workspace management, and terminal sessions, while a React frontend (built with Node.js and pnpm) provides the kanban board and review interface. The two communicate via HTTP APIs, with optional relay/tunnel mode for remote server deployments.
The killer feature is workspace isolation through git worktrees. When you create a workspace from a kanban issue, the backend creates a new git worktree—essentially a separate working directory linked to a dedicated branch. Each workspace gets its own terminal session where you can invoke any coding agent from the supported list (Claude Code, Codex, Gemini CLI, GitHub Copilot, Amp, Cursor, OpenCode, Droid, CCR, and Qwen Code—10+ agents total). Because worktrees share the same git repository but have separate working directories, multiple agents can work on different features simultaneously without stepping on each other’s toes. No branch juggling, no stash conflicts, no “wait, which terminal was I using for that feature?”
Here’s what the workflow looks like from the command line:
npx vibe-kanbanThat single command starts the backend server and opens the UI. The simplicity is deceptive—under the hood, the Rust backend is spawning a database for persistence, initializing git repository watchers, and setting up HTTP endpoints for workspace management. The frontend connects and presents your kanban board. Create an issue, click “Create Workspace,” and Vibe Kanban provisions a new worktree, checks out a branch, and spins up a terminal session.
The agent abstraction is deliberately thin. Vibe Kanban doesn’t try to wrap or proxy agent commands—it assumes you’ve already authenticated with your preferred agent and simply provides a terminal where you can invoke it normally. Want to use Claude Code? Run claude-code "implement user authentication" in the workspace terminal. Prefer Codex? Run codex "add OAuth support". The backend captures output and watches for file changes, but the agent itself is just a subprocess. This keeps Vibe Kanban agent-agnostic and avoids the vendor lock-in that plagues tools built around specific AI providers.
The diff review interface is where the integration shines. When your agent modifies files, the frontend automatically displays side-by-side diffs with inline commenting. You can leave feedback directly in the UI, then copy it back to the agent’s terminal or use it to refine your next prompt. No switching to git diff or opening GitHub. The built-in browser preview is equally seamless—it includes full devtools, inspect mode, and device emulation, so you can test responsive layouts without juggling Chrome tabs.
Once you’re satisfied with the changes, Vibe Kanban can generate a PR description using AI, create the pull request via GitHub’s API, and update the kanban issue status. The entire loop—plan, execute, review, merge—happens in one interface. The reduction in context switching is substantial enough that it genuinely changes how planning and review feel.
The relay/tunnel architecture deserves mention for self-hosting scenarios. By setting VK_TUNNEL and VK_SHARED_RELAY_API_BASE, you can run the backend on a remote server while keeping the frontend accessible via a custom domain. The relay acts as a reverse proxy for tunnel-mode connections, routing requests to your backend without exposing it directly. It’s a thoughtful design for teams that want centralized agent management without the security exposure of a public-facing server.
Gotcha
The onboarding friction is real. The README guides you to “make sure you have authenticated with your favourite coding agent” before running, which means you need to install and configure Claude Code, Codex, or whichever agent you’re using before Vibe Kanban becomes useful. There’s no guided setup, no OAuth flow, no configuration wizard. If your agent isn’t already in your PATH and authenticated, you’re debugging that before you can even try the core workflow. For teams evaluating multiple agents, this means separate authentication for each one.
The git worktree dependency is both a strength and a liability. Worktrees are powerful but not universally loved—they can behave unexpectedly with certain git configurations, submodules, or very large repositories. If your repo has unusual hooks or relies on specific working directory assumptions, worktrees might break things. The README mentions DISABLE_WORKTREE_CLEANUP for debugging, which suggests the team has encountered edge cases in worktree management.
The contribution model is restrictive. The README explicitly says “Please do not open PRs without first discussing your proposal with the team.” While this keeps the codebase coherent, it limits community-driven improvements and signals that this is more of a product than a community project. If you hit a bug or want a feature, you’re dependent on the core team’s priorities. The 24,000+ stars suggest demand, but the gatekept contribution process means that momentum might not translate to rapid iteration.
Verdict
Use Vibe Kanban if you’re already deep into AI-assisted development and spend significant time juggling agents, diffs, and kanban boards. The workspace isolation and unified review interface genuinely compress the plan-execute-review cycle, and the multi-agent support (10+ coding agents including Claude Code, Codex, Gemini CLI, GitHub Copilot, Amp, Cursor, OpenCode, Droid, CCR, and Qwen Code) means you’re not locked into a single vendor’s ecosystem. It’s especially valuable for teams running multiple agents in parallel or individuals managing several AI-driven projects simultaneously. The self-hosting options make it viable for companies with strict data residency requirements. Skip it if you rarely use coding agents or prefer tools that embed AI directly into your existing workflow without requiring a separate orchestration layer. Also skip if you need traditional project management features like team permissions, sprints, or roadmapping—Vibe Kanban’s kanban is deliberately minimal, focused on developer throughput rather than stakeholder reporting. Finally, if you want a community-driven project where you can easily contribute fixes and features, the restrictive PR policy might frustrate you. This is a tool for practitioners who’ve already decided agents are central to their workflow and need better infrastructure around them, not for teams still experimenting with whether AI coding assistants are worth the investment.
