What It Does
Vibe Kanban is a local-first, open-source web application that provides an orchestration shell for AI coding agents. Rather than being an agent itself, it sits above agents like Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor, Amp, and OpenCode — wrapping them with a kanban planning board, isolated per-task workspaces using git worktrees, inline diff review with commenting, and an embedded browser with developer tools for live application preview.
The core workflow is “describe the work, review the diff, ship it.” Developers create issues on a kanban board, assign them to an agent, each task runs in its own git worktree branch, and the resulting diff is reviewed and merged to a PR — all within a single interface. The stack is Rust backend (50% of codebase) + TypeScript/React frontend (46%), launched via npx vibe-kanban.
Key Features
- Kanban board for task planning: Create, prioritize, and assign issues; each issue becomes an agent workspace with its own branch
- Git worktree isolation: Each workspace gets a dedicated git branch and working directory, enabling genuine parallel agent execution without file-system conflicts
- Inline diff review: Review agent-generated changes with inline comments without leaving the UI
- Embedded browser with DevTools: Built-in browser supporting inspect mode, device emulation, and developer tools for live application preview during development
- 10+ agent support: Claude Code, Codex, Gemini CLI, GitHub Copilot, Amp, Cursor, OpenCode, Droid, CCR, Qwen Code — abstracted as subprocess commands
- Auto-generated PR descriptions: AI-generated pull request descriptions for agent-completed tasks
- Self-hosting options: Docker support, SSH-based remote access, VK_TUNNEL relay mode for cloud/remote scenarios
- Single-command startup:
npx vibe-kanbanwith no persistent installation required - Optional analytics: PostHog integration disabled when API keys are absent
Use Cases
- Parallel agent execution: Teams or solo developers who want to run multiple coding agents simultaneously on different tasks, each in isolated branches, without context-switching between terminals
- Agent output review workflow: Engineering leads who want a structured interface for reviewing AI-generated diffs before merging, replacing ad-hoc terminal sessions
- Agentic sprint planning: Decomposing a feature into sub-tasks, assigning each to an agent, and managing the review/merge lifecycle from one UI
- Local development with preview: Frontend/fullstack developers who want to see live results of agent-generated changes in an embedded browser alongside the diff
Adoption Level Analysis
Small teams (<20 engineers): Fits well for teams already using AI coding agents and wanting to organize parallel work. Single-command startup, free, and open-source. The lightweight subprocess-based agent abstraction works without infrastructure overhead. However, rough edges (358 open issues, worktree lifecycle bugs, integration gaps across agents) require tolerance for early-adopter friction. The Claude Code settings override bug is a concern for teams using hook-based guardrails.
Medium orgs (20-200 engineers): Potentially fits for teams with established agentic workflows who need a lightweight coordination layer. The git worktree model scales reasonably for parallel task execution. However, no multi-user support, no RBAC, no audit logging, and no enterprise authentication (Entra ID requested but not implemented) limit organizational deployment. Best used as a per-developer local tool rather than a shared team infrastructure component today.
Enterprise (200+ engineers): Does not fit today. No enterprise auth, no governance features, no compliance tooling, and no SLA. The settings override issue with Claude Code and documented integration instability across agents make it unsuitable for environments with strict guardrails. Enterprises should evaluate OpenHands or dedicated agent orchestration platforms instead.
Alternatives
| Alternative | Key Difference | Prefer when… |
|---|---|---|
| OpenHands | Builds its own Docker sandbox runtime; more isolation and agent control | You need sandboxed execution with Docker isolation, not just subprocess orchestration |
| Beads | Dependency-aware task graph for agent memory; CLI tool, not a UI shell | You want persistent agent memory and dependency tracking across sessions, not a visual kanban |
| Claude Code (direct) | Single agent, CLI-native, no orchestration overhead | You work with one agent at a time and want the most polished single-agent experience |
| GitHub Issues + Claude Code | Native code hosting integration, PR workflow, no extra UI | You want agents operating directly in GitHub’s workflow without a separate orchestration layer |
| Optio | K8s-native workflow orchestration for agents | You need production-grade workflow orchestration with enterprise infrastructure integration |
Evidence & Sources
- GitHub: BloopAI/vibe-kanban — source code, 24.8k stars, 2.5k forks, 358 open issues
- GitHub Issues tracker — primary source for known bugs and limitations including settings override, workspace disappearance, and agent integration gaps
Notes & Caveats
- Claude Code settings override is a security concern. Documented GitHub issue confirms that project-level
.claude/settings.jsonhooks (pre-commit, PostToolUse) are silently overridden by Vibe Kanban’s SDK Initialize message. Teams relying on hooks for guardrails (secret detection, linting enforcement, tool permissions) must verify this is fixed before using in any security-sensitive environment. - Worktree lifecycle is unstable. Workspaces can disappear from kanban cards after cleanup, and the
DISABLE_WORKTREE_CLEANUPdebug flag signals that cleanup edge cases are not fully resolved. This can result in orphaned worktrees or lost workspace state. - Integration depth varies across agents. The subprocess abstraction that enables multi-agent support also limits how deeply agent-specific features (Claude Code hooks, Codex async delegation, Copilot CLI slash commands) are surfaced. Integration quality is uneven — Claude Code and likely Codex are most tested; others may have gaps.
- No multi-user support. The tool is designed for a single developer’s machine. There is no shared workspace, team access control, or concurrent user support.
- BloopAI pivot risk. The company previously built Bloop (code search), which appears to have wound down or pivoted. Vibe Kanban represents a new direction. The sustainability of a small company maintaining a free open-source tool without a clear monetization model warrants monitoring.
- 358 open issues at launch traction. The high issue count relative to project age (24.8k stars, 2.5k forks) reflects rapid adoption outpacing engineering bandwidth — a common pattern in viral developer tools. Expect continued instability alongside rapid iteration.
- Rust + React monorepo stack. Development requires Rust (latest stable), Node.js ≥20, pnpm ≥8, cargo-watch, and sqlx-cli. This is higher friction for contributors than pure JS/TS projects. The Rust backend provides performance benefits but narrows the contributor pool.