Vibe Kanban: A Genuinely Useful AI Agent Orchestrator — That's Now Shutting Down
25,000 stars in under a year. That's the kind of traction that makes you stop and pay attention. Vibe Kanban landed in a moment when every developer I know is simultaneously excited about AI coding agents and frustrated by how chaotic running them actually is. It clearly hit a nerve. But before you go npx vibe-kanban and start wiring this into your workflow, there's something you need to know upfront: Bloop, the company that built this, is shutting down. The latest release even added a data export feature specifically so users can get their data out before the lights go off. I'll get into what that means for you, but let's first talk about what this thing actually does — because it's genuinely interesting work.
What It Actually Does
Vibe Kanban is a local-first project management layer designed specifically for workflows where AI agents are doing most of the coding. Think of it as a task board that knows about Claude Code, Codex, Gemini CLI, and a dozen others — and can actually spin them up, point them at a branch, and let you review what they did, all from one UI.
The core loop is: create an issue on a kanban board, open a "workspace" for that issue, pick your agent, let it run, review the diff with inline comments, and merge. Each workspace gets its own git worktree, terminal, and a built-in browser with devtools. You're not just running an agent in a terminal and hoping for the best — you have a structured way to track what was asked, what was done, and what needs review.
The backend is Rust (fast, low overhead), the frontend is a React/TypeScript SvelteKit-style monorepo managed with pnpm, and the whole thing ships as a single npx command. The architecture is actually pretty clean for what it's doing.
Why This Exists and Why It Resonated
Here's the gap it fills: if you're using Claude Code or Codex seriously, you're probably running multiple tasks in parallel across different branches. You've got one agent refactoring auth, another writing tests, a third fixing that CSS bug. Keeping track of all of this in your head — or worse, in a separate Notion board — is genuinely painful.
Vibe Kanban puts the task management and the agent execution in the same place. That's not a revolutionary idea, but nobody had actually built it well before this. The timing also aligned perfectly with Claude Code going mainstream and developers realizing that "run one agent at a time and babysit it" doesn't scale.
The 25k stars reflect real developer frustration finding a real solution. The community around it was active — 506 open issues (which is a lot, but also signals genuine usage), a Discord, and a contribution model that required discussion before PRs. That last part is a green flag for me: it suggests the core team was thinking about architecture, not just merging everything.
Features Worth Calling Out
Multi-agent support across 10+ tools. Claude Code, Codex, Gemini CLI, GitHub Copilot, Amp, Cursor, OpenCode — the list is long and they actually maintain it. Switching between agents on different tasks isn't an afterthought, it's the main use case.
Git worktree isolation per workspace. Each task gets its own branch and worktree automatically. This is the right call. It means parallel agent runs don't stomp on each other, and you always have a clean diff to review. Anyone who's tried to run two Claude Code sessions in the same repo directory knows why this matters.
Built-in diff review with inline comments. You can leave comments on the diff that get sent back to the agent. This is the feature that actually closes the loop — instead of copy-pasting diffs into a chat window and re-explaining context, the feedback stays attached to the task. It's the right UX for this workflow.
Integrated browser preview with devtools. This is a nice touch. You can preview what the agent built, inspect elements, toggle device emulation — without leaving the tool. It's not a replacement for your actual browser, but for quick sanity checks it removes friction.
Self-hosting support. Docker deployment, reverse proxy configuration, remote SSH integration — the docs cover it. For teams who don't want to run everything locally, this was a real option. The VK_ALLOWED_ORIGINS configuration is a bit of a footgun if you skip it behind a proxy, but it's documented.
Who Should Use This (Right Now)
Honestly? The answer to this has changed because of the shutdown.
If you want to evaluate the concept: Spin it up, play with it, understand the workflow. It's a great reference for what a well-designed AI agent orchestration UI looks like. The code is Apache-2.0, so you can read it, fork it, learn from it.
If you're a team considering building something similar: This is worth studying. The Rust backend with a React frontend, the worktree-per-workspace model, the MCP server integration — there are real architectural decisions here that are worth understanding before you roll your own.
If you want to use it in production today: I'd be cautious. The cloud service is going away. The local-only mode still works, and the npm package is published, but you're adopting software from a company that's winding down. That's not a dealbreaker for open source tools with strong communities, but the commit velocity will drop and the 506 open issues aren't going to get triaged.
Concerns and Limitations — Being Honest
Let me be direct about the things that would give me pause even if Bloop weren't shutting down.
506 open issues is a lot for a tool this young. The repo is less than a year old. That issue count suggests either very active usage (good) or a backlog that the team was struggling to keep up with (less good). Probably both. Some of those issues are going to be real bugs that never get fixed now.
The "10X" marketing is the kind of thing I reflexively distrust. The tool is useful. It doesn't need the hype. When I see that in a description, I mentally discount the marketing and look harder at the actual functionality. The actual functionality holds up — but lead with that.
The cloud/local split adds complexity. There are environment variables for relay APIs, tunnel modes, shared API bases — the architecture clearly evolved from a pure local tool toward a cloud product, and now the cloud side is going away. The local path should still work, but some of the configuration surface area is there to support a product that won't exist.
Dependency on agent CLIs being installed and authenticated. Vibe Kanban doesn't replace your agent setup — it orchestrates on top of it. If Claude Code breaks an API or changes its auth flow, you'll feel it here. That's not a criticism of Vibe Kanban specifically, it's just the reality of building on top of fast-moving tools.
The "We're hiring" banner in the README is still there. Small detail, but it's the kind of thing that doesn't get updated when a company is shutting down. It makes me wonder what else in the docs is stale.
The Shutdown Context
I want to spend a moment on this because it's the elephant in the room. The latest release note literally says "add cloud data export feature for service shutdown." The team shipped a feature specifically so users could get their data out. That's actually a responsible thing to do — it's more than a lot of companies do. But it also means the people who built this deeply are moving on.
The codebase is Apache-2.0. Someone could fork it and maintain it. The npm package is published. The architecture is solid enough that a motivated team could pick this up. But right now, that hasn't happened publicly. If you're evaluating this for a team, you should factor in that you might be inheriting maintenance responsibility.
Verdict
Vibe Kanban is genuinely good software that solved a real problem at the right time. The architecture is thoughtful, the core workflow is well-designed, and the 25k stars reflect actual developer interest rather than hype. If you're building AI-assisted development workflows, understanding how this tool works is worth your time.
But I can't recommend adopting it as a production dependency right now. The company is shutting down, the issue backlog is large, and the cloud features are going away. If you want to run it locally for personal projects, go for it — it works and it's useful. If you're evaluating tools for a team, look at what's being built on top of similar ideas, or consider whether this codebase is worth forking and maintaining.
The real value here might be the blueprint. Vibe Kanban shows what a well-designed AI agent orchestration layer looks like: git worktree isolation, structured task management, inline review feedback, multi-agent support. Whatever tool you end up using — or building — these are the right ideas.