⚡ VibeBoard

The kanban board that runs your AI coding agents.

Move a card to In Progress → VibeBoard spawns Claude Code, OpenCode, or Codex in your repo, hands it the task, and streams its progress back to the board in real time. You and your agents drive the same board - through the UI and through MCP - at the same time.

npx @zanuartri/vibeboard

---

Why VibeBoard

AI coding agents are great at doing the work - but you still need somewhere to plan it, track it, and watch it happen. Chat windows aren't that place.

VibeBoard makes a kanban board the control surface for your agents:

• You plan on the board like normal - columns, cards, priorities, tags, due dates.
• Move a card to "In Progress" and the assigned agent launches in your project directory with the card as its brief.
• The agent reports back through MCP tools - posting checkpoints, moving its own card, opening a PR - so the board is always the source of truth.

It's local-first, self-hosted, and dependency-free to run. One command, no account, no cloud.

Demo

Screenshots / GIF coming soon. Run npx @zanuartri/vibeboard and open http://localhost:7341 to try it in under a minute.

  Backlog            In Progress              Review            Done
 ┌──────────┐      ┌────────────────┐      ┌──────────┐     ┌──────────┐
 │ Fix login│      │ Add OAuth   ⚡ │      │ Refactor │     │ Landing  │
 │ Add tests│  →   │ ● agent running│  →   │ db layer │  →  │ page  ✓ │
 └──────────┘      └────────────────┘      └──────────┘     └──────────┘
                     ↑ card moved here → Claude Code spawned in ./my-repo

Quick start

# Run instantly (no install)
npx @zanuartri/vibeboard

# …or install the CLI globally
npm install -g @zanuartri/vibeboard
vibeboard

Then open http://localhost:7341, create a workspace pointing at a project directory, and start adding cards.

Prefer to hack on it?

git clone https://github.com/zanuartri/vibeboard.git
cd vibeboard && npm install && npm start

Requirements: Node.js 18+. VibeBoard uses better-sqlite3, which ships prebuilt binaries for common platforms - most installs need no compiler. On an unsupported platform/Node it builds from source (needs standard build tools).

How it works

   ┌─────────────┐         move card to "In Progress"         ┌──────────────┐
   │   You (UI)  │ ───────────────────────────────────────►   │   VibeBoard  │
   └─────────────┘                                             │   server     │
          ▲                                                    └──────┬───────┘
          │ live SSE updates (notes, status, moves)                   │ spawns
          │                                                           ▼
   ┌──────┴───────┐         MCP tools (get_board, move_card,   ┌──────────────┐
   │  The board   │ ◄────── add_card_note, complete_card …) ── │  Agent in    │
   │ (shared)     │                                            │  your repo   │
   └──────────────┘                                            └──────────────┘

1. Create a workspace - link a project directory. VibeBoard detects the git repo and can isolate each agent in its own git worktree / branch.
2. Add a card - title, description, tags, priority, due date, and an agent (Claude Code, OpenCode, or Codex CLI).
3. Move it to In Progress - the agent spawns in your repo with full card context and starts working autonomously.
4. Watch it live - the agent posts checkpoints, you stream its output, and the card shows run status, duration, and cost when it finishes.
5. Ship it - review the diff in the card, then Merge or Create PR right from the board.

Features

• ⚡ Auto-spawn agents - moving a card to In Progress launches its assigned agent in your project directory. No copy-pasting prompts.
• 🔄 Bidirectional MCP - you and the agent control the same board in real time. Agents read state, post progress, move cards, and mark work done.
• 🧵 Git worktree isolation - each agent works on its own branch; review the diff and merge or open a PR from the card.
• 📊 Run visibility - live output stream, agent checkpoints, plus exit code, duration, and token/cost per run.
• 🚦 Flow controls - WIP limits per column, card dependencies (blocked_by), and a concurrency cap with an automatic agent queue.
• 🗂️ Multiple workspaces - one board per project, switch instantly.
• 🎨 ClickUp-inspired UI - clean white surfaces, colorful column accents, vibrant workspace icons, skeleton loading, full light & dark theme.
• 🔒 Local-first - SQLite on your machine. No cloud, no account, no tracking.

Connect your agents

Open the in-app MCP Setup dialog to auto-configure each agent's global MCP config in one click - or wire them up manually:

Claude Code

claude mcp add -s user vibeboard -- node /path/to/vibeboard/mcp-server/index.js

(The repo's .claude/mcp.json is also auto-detected when working inside it.)

OpenCode

{
  "mcp": {
    "vibeboard": { "type": "local", "command": ["node", "/path/to/vibeboard/mcp-server/index.js"] }
  }
}

Codex CLI

Add a vibeboard entry under mcpServers in ~/.codex/config.json:

{ "mcpServers": { "vibeboard": { "command": "node", "args": ["/path/to/vibeboard/mcp-server/index.js"] } } }

MCP tools agents can call

get_board · get_column · list_cards · search_cards · list_workspaces · create_workspace · switch_workspace · set_workspace · create_card · update_card · move_card · complete_card · delete_card · add_card_note · get_card_notes · get_agent_status · list_models · refresh_models · list_templates · create_template

get_board supports columnsOnly, excludeLogs, and columnTitle filters to reduce payload size. search_cards and list_cards support limit/offset pagination.

Configuration

All optional - set as environment variables (a .env file works too):

Variable	Default	Description
PORT	7341	HTTP/UI port.
VB_HOST	127.0.0.1	Bind address. Loopback only by default.
VB_TOKEN	(random)	Shared token required for remote clients when VB_HOST is non-loopback. Auto-generated and printed at startup if unset.
VB_MAX_AGENTS	3	Max agents running at once; extra spawns are queued.
AGENT_TIMEOUT_MS	1800000	Per-agent run timeout (30 min).

⚠️ Network exposure is opt-in and powerful. Agents run with elevated permissions in your project directories. When VB_HOST is non-loopback, VibeBoard requires a token for off-machine requests - open it via http://<host>:7341/?token=<token>. Only expose it on a trusted network. See the security notes before exposing anything.

Data & storage

Everything lives in a local SQLite database:

• Windows: %APPDATA%\vibeboard\vibeboard.db
• macOS: ~/Library/Application Support/vibeboard/vibeboard.db
• Linux: ~/.local/share/vibeboard/vibeboard.db

Export/import any workspace as JSON from its settings.

Security

VibeBoard spawns coding agents with permissions skipped so they can work unattended - so only point it at projects you trust, and keep it on loopback unless you've read the network notes above. Mutating requests are guarded against CSRF / DNS-rebinding, and agents are terminated on server shutdown. It's early software (pre-1.0); review before using on sensitive code.

Architecture

No build step. Plain Node.js (CommonJS) server, plain HTML/CSS/vanilla-JS client.

mcp-server/   Node server - index.js wires http-routes.js (REST+SSE),
              mcp-tools.js (MCP), agent.js (spawn/queue), worktree.js,
              db.js (SQLite), auth.js, events.js, models.js, migrate.js
public/       index.html + styles.css + ordered js/*.js classic scripts
              (no bundler), and landing.html

See DESIGN.md for the UI design system and CLAUDE.md / AGENTS.md for agent-facing context.

Contributing

Contributions are welcome - see CONTRIBUTING.md. In short: fork → branch → npm test → PR. Please keep Windows/macOS/Linux compatibility, don't break MCP tool signatures, and stay backward-compatible with existing workspaces.

See CHANGELOG.md for release history.

License

MIT © VibeBoard contributors