How it stacks up.
The honest version. Three tools people actually put next to Kungfu Kanban when they're deciding how to run coding agents: Vibe Kanban (the other kanban board for agents), and Conductor / Crystal (now Nimbalyst) — worktree-per-task session managers people evaluate for the same job. Every claim below is something you can check yourself in each project's docs or repo.
At a glance
| Thing | Kungfu Kanban | Vibe Kanban | Conductor | Crystal / Nimbalyst |
|---|---|---|---|---|
| Interface | Kanban columns: Backlog → Queued → Running → Review → Done | Kanban board | One tab per task workspace | One tab per session |
| Agent backends | Claude Code only | 10+ (Claude Code, Codex, Gemini CLI, Copilot, Cursor, Amp, OpenCode, and more) | Claude Code, Codex, Cursor | Claude Code, Codex |
| Billing | Claude subscription only — ANTHROPIC_API_KEY stripped from every run |
whatever the chosen backend bills — API keys for most non-Claude agents | whatever the chosen backend bills | whatever the chosen backend bills |
| Task import | Markdown paste/drop, open GitHub issues, LLM drafting | manual card creation | manual workspace creation | manual session creation |
| Dependencies | “Runs after” chains gated on the PR actually being merged, not just card status | board can filter to unblocked work | no queue/dependency model — each workspace stands alone | no queue/dependency model — each session stands alone |
| Reviewer | LLM manager (the Sensei) reads the real PR diff, merges green PRs itself within its autonomy setting, sends targeted follow-ups instead of blind retries | you review diffs and leave inline comments in the UI | you review via GitHub/Linear; the agent can respond to review comments | you review diffs in-app |
| Data | plain JSON files in data/, nothing else |
SQLite database managed by a compiled Rust server | GitHub/Linear as system of record | local SQLite (Electron) |
| Install footprint | one npm dependency (express) |
compiled Rust binary + SQLite + React, shipped via npx |
native Mac app | Electron app |
| License / status | MIT, self-hosted, free | MIT — sunsetting as a company product, continuing community-maintained | free today; Series A-funded company, paid collaboration features planned | MIT, but development moved to a proprietary successor (Nimbalyst) |
| Phone access | token gate + Tailscale serve, installs as a PWA | not part of the product | Mac only | desktop only |
vs Vibe Kanban
Vibe Kanban is the closest thing to a direct competitor: an open-source kanban board (BloopAI) purpose-built for orchestrating coding agents, with real strengths Kungfu Kanban doesn't have — 10+ supported agent backends (Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor, Amp, OpenCode, and more), an in-app browser preview with devtools and device emulation, and virtualized diff review built for large changesets. If you want one board to run several different coding agents side by side, it's the better fit.
The split is billing and footprint. Vibe Kanban's cost model follows whichever backend a
card uses — plenty of those bill per API token. Kungfu Kanban only drives the Claude Code
CLI, and the runner deletes ANTHROPIC_API_KEY from every card's environment so
a run can never quietly fall back to pay-per-token; everything runs against your existing
Claude subscription. Vibe Kanban ships a compiled Rust server plus a SQLite database via
npx; Kungfu Kanban is one Express process with one dependency and plain JSON
files you can read with a text editor. And where Vibe Kanban's review step is a human reading
diffs in the UI, Kungfu Kanban's Sensei is an LLM manager that reads the actual PR diff itself
and can merge it when every CI check is green, within whatever autonomy level you've set.
Worth knowing: Vibe Kanban is sunsetting as a company product and continuing as community-maintained open source — a fact worth weighing if roadmap continuity matters to you.
vs Conductor
Conductor is a well-built Mac app for running Claude Code, Codex, or Cursor in parallel: every new task gets its own git-worktree workspace, its own branch, its own agent, and connects with GitHub and Linear so the agent can open PRs, read diffs, comment, and respond to review feedback. If your team already lives in GitHub/Linear and wants a slick native Mac experience, that's a genuine strength.
It isn't a kanban board, though — there's no backlog/queued/running/review/done flow, no markdown or GitHub-issue import, and no dependency chains between tasks the way “Runs after” works in Kungfu Kanban. It's also a funded company product (a $22M Series A as of mid-2026) that's free today with paid collaboration features on the roadmap, versus Kungfu Kanban's MIT license and plain-JSON local state with nothing to migrate if a company changes direction.
vs Crystal / Nimbalyst
Crystal (MIT-licensed, Electron) runs multiple Claude Code or Codex sessions in parallel git worktrees, with session persistence, diff/change tracking, and desktop notifications — a clean way to compare a few approaches to the same task side by side. As of February 2026 the project was renamed and its active development moved to a proprietary successor, Nimbalyst; the original Crystal code still runs but isn't the maintained version going forward.
Like Conductor, it's a session manager rather than a task board: one tab per session, no queue, no dependency chains, no markdown/issue import, no manager reviewing and merging work on your behalf. Kungfu Kanban's Sensei, merge-gated dependency chains, and scheduled/recurring cards don't have an equivalent here.
Which one to pick
- Vibe Kanban — you want one board driving several different agent backends, not just Claude Code, and you're fine with each backend's own billing.
- Conductor — you're on a Mac, you want tight GitHub/Linear integration, and a workspace-per-task tab layout suits how your team already works.
- Crystal / Nimbalyst — you mainly want to run a few parallel attempts at the same task and compare them, not manage an ongoing backlog.
- Kungfu Kanban — you want everything to run on the Claude subscription you already pay for with no separate API bill possible, a queue that imports from markdown or GitHub issues, dependency chains gated on real merges, and a manager that reads the actual diff before it ships anything — all as one Express server and a folder of JSON you fully own.