Vibe Kanban

Vibe Kanban was a good idea and it is over as a product. Bloop, the company, announced on 10 April 2026 that it was closing, and what remains is a community-maintained open-source board that still drives Claude Code, Codex, Gemini CLI and a long list of others in local worktrees. The daily experience is fine; the future is a volunteer's weekend.

Do not build a habit on it. If you want a board over parallel agents, pick Agent Orchestrator, and if you want a desktop app with a company behind it, pick Superset.

The failure mode is now the issue tracker. There is no company reading it, so a broken agent adapter waits for a volunteer with the same problem. The board runs agents directly on the host with no container layer, so a runaway process in one card is a runaway process on your laptop, and with no vendor there is nobody to ship the fix on a schedule.

Pin the version you have. What it does right: pull requests are created from the board rather than merged by it, so integration stays in code review where a human sees it.

The design partitions by card. 1. Each card owns a git worktree, a terminal and a dev server, so parallel tasks never share a tree or a port. 2. A built-in browser previews the running app beside the diff. 3. The board exposes an MCP server, so an agent can create and move cards, closing the loop from supervisor to supervised.

No benchmark was ever published. The observation: the architecture was ahead of its business model, which is the usual order and the usual outcome; the code survives the company, the roadmap does not.

The post-mortem is on the blog and it is unusually candid: the majority of users were on the free tier, and the team could not find a business model it could get excited about. Refunds were issued for invoices paid in the previous thirty days and subscriptions terminated. That is the honourable version of running out of road.

The lesson for every free-plus-hosted orchestrator on this board: a worktree manager has no data moat and no switching cost, so hosted features must be worth paying for on day one. Position: none; there is no company to have a position on.

The demo still runs from npx vibe-kanban. Procurement has nothing to procure: the remote services that held organisations, projects and comments were shut down thirty days after the announcement and the app went fully local. No SSO was ever the question; now there is no counterparty, no contract, no support line and no roadmap.

Sixty engineers on an unowned tool is sixty unpatched installs. It never ran in CI. Onboarding is trivial and irrelevant. Not yet, and unless a maintainer organisation forms, not later either.

Apache-2.0, so the company's exit changed my rights not at all. The board serves MCP, meaning I can script card creation from any agent that speaks it, and the whole thing runs from a single npx command with no account, which is what the shutdown reduced it to anyway. Twenty-eight thousand stars is a lot of people who could send a patch.

No MCP client and no model settings, because it manages agents rather than being one. The fork is the project now; that is the honest state of it, and for my purposes it works.