A web-based command center for managing Claude Code CLI sessions through your browser. Spawn, monitor, and interact with multiple sessions from any device.
Built to run on a home server (e.g. Mac Mini) and accessed remotely over Tailscale.
- Multi-session management — create, monitor, and kill sessions from a dashboard
- Browser-based terminal — full xterm.js terminal with color, resize, and clickable links
- Git worktree workflow — clone repos, create feature worktrees, and assign tasks to Claude
- Dashboard grouping — sessions are grouped by repository for easy navigation
- Durable sessions via tmux — sessions survive server restarts; SSH in and
tmux attachas fallback - Dual transport — WebSocket with automatic SSE fallback for slow/unreliable networks
- Status detection — polls tmux pane content to show session state (idle, running, waiting)
- Quick actions — approve, reject, or send custom input to sessions without opening the terminal
- Push notifications — get notified on your phone when Claude needs input (via ntfy)
- Simple auth — passphrase + signed cookie (designed for use behind Tailscale)
- Mobile-friendly — responsive layout, works on iPhone Safari
Click a session card to open a full browser-based terminal powered by xterm.js.
Create sessions in three ways:
Directory — point Claude at any directory on your server.
Git Clone — clone a repo into a worktree-friendly structure (repo/repo-main/) and start Claude in it.
Worktree — select a previously cloned repo, name a feature branch, and optionally describe the task. Claude starts in a new git worktree and receives your task description as its initial prompt.
Browser (MacBook / iPhone over Tailscale)
│
├── HTTP: REST API + SSE fallback
└── WebSocket: live terminal
│
Server (Node.js + Express)
│
├── SQLite: session metadata
├── node-pty: terminal bridge
└── tmux: session backend (durable)
- Node.js 20+ (tested with 23)
- tmux (
brew install tmux) - Claude Code CLI installed and authenticated
- Tailscale (for remote access)
git clone https://github.com/Afstkla/claude-command-center.git
cd claude-command-center
# Install dependencies
npm install
cd frontend && npm install && cd ..
# Configure
cp .env.example .env
# Edit .env — set AUTH_PASSPHRASE and AUTH_SECRET| Variable | Description | Default |
|---|---|---|
PORT |
Server port | 3100 |
AUTH_PASSPHRASE |
Login passphrase | (required) |
AUTH_SECRET |
JWT signing key (min 32 chars) | (required) |
COOKIE_MAX_AGE_HOURS |
Auth cookie expiry | 24 |
npm run dev
# Backend: http://localhost:3100
# Frontend: http://localhost:5173 (proxies API to backend)npm run build
npm startUse pm2 to keep the server running and auto-restart on boot:
npm install -g pm2
# Start the server
pm2 start npm --name command-center --cwd /path/to/claude-command-center -- start
# Save the process list and enable startup on boot
pm2 save
pm2 startup # follow the printed instructions (requires sudo)Important: Do not run
pm2 startfrom inside a Claude Code session. Claude Code sets aCLAUDECODEenvironment variable that pm2 captures, which prevents spawned sessions from starting. If you already did this, fix it with:pm2 delete command-center env -u CLAUDECODE -u CLAUDE_CODE_ENTRYPOINT pm2 start npm --name command-center --cwd /path/to/claude-command-center -- start pm2 save
Since this is designed to run on a headless server accessed remotely, network stability matters. Here are common issues and fixes discovered while running on a Mac Mini over Tailscale.
macOS defaults to sleeping after a few minutes of inactivity, even on a Mac Mini. Each sleep cycle kills all TCP connections (WebSocket, SSE, API streams). Disable it:
sudo pmset -a sleep 0 disksleep 0 powernap 0Alternatively, keep caffeinate running via pm2:
pm2 start caffeinate --name no-sleep -- -s
pm2 saveIf both Ethernet and Wi-Fi are active on the same subnet, macOS may route traffic through either interface unpredictably. This causes long-lived connections (WebSockets, SSE) to break when routes flip. Disable Wi-Fi if you have a wired connection:
networksetup -setairportpower en1 offThe terminal uses several mechanisms to handle unstable connections:
- Server-side ping/pong — the WebSocket server pings clients every 15s and terminates unresponsive connections, preventing zombie sockets
- SSE keepalive — periodic comments prevent proxies and VPN tunnels from dropping idle SSE connections
- Automatic reconnect — the frontend reconnects indefinitely with exponential backoff (up to 30s), instead of giving up after a few attempts
- Client-side heartbeat — if no data arrives within 25s, the client proactively reconnects rather than waiting for a TCP timeout
- Dashboard offline banner — shows a "Connection lost" indicator when the API is unreachable, clears automatically on recovery
- Visibility-aware refresh — the dashboard refetches immediately when the browser tab regains focus or the device comes back online
A monitoring script is included for debugging connectivity issues:
pm2 start scripts/network-monitor.sh --name network-monitor --interpreter bashLogs to logs/network-monitor.log every 30s with: interface status, default route, ping times, DNS resolution, TCP/HTTPS latency to the Anthropic API, and Tailscale state. Entries are flagged as PROBLEM with extra detail when checks fail.
- Install Tailscale on your server and client devices
- Run
tailscale upon both - Find your server's Tailscale IP:
tailscale ip -4 - Access the command center at
http://<tailscale-ip>:3100
No HTTPS needed — Tailscale encrypts all traffic via WireGuard. No port forwarding, no public exposure.
- Zero-config VPN: devices authenticate via your identity provider
- WireGuard encryption: all traffic is end-to-end encrypted
- No open ports: your server is never exposed to the public internet
- Works everywhere: train wifi, coffee shop, mobile data
- Open
http://<tailscale-ip>:3100and enter your passphrase - Click New Session and choose a template:
- Directory — pick a folder and start Claude in it
- Git Clone — clone a repo (creates
repo/repo-main/structure for worktree support) - Worktree — create a feature branch worktree from an existing repo, with an optional task prompt
- Click a session card to open the terminal
- Use quick action buttons on session cards to approve, reject, or send input without opening the terminal
The worktree workflow lets you run multiple Claude sessions on the same repo without conflicts:
- Clone a repo using the Git Clone template — this creates a
repo/repo-main/directory structure - Create worktrees using the Worktree template — each gets its own branch (
feat/<name>) and directory - Assign tasks via the Task Description field — Claude receives it as its first prompt after startup
- Clean up — killing a worktree session automatically removes the git worktree
Sessions from the same repo are grouped together on the dashboard.
Sessions run inside tmux. If the server crashes or restarts:
- Sessions keep running in tmux
- The server reconnects to them on startup
- You can always SSH in and run
tmux ls/tmux attach -t cc-<id>
Get notified on your phone when Claude needs tool approval or wants your input. Uses ntfy for push delivery.
Add ntfy config to your .env:
NTFY_ENABLED=true
NTFY_URL=https://ntfy.sh # or your self-hosted instance
NTFY_TOPIC=your-secret-topic
NTFY_AUTH_TOKEN=your-token # optional, for private topics
BASE_URL=http://<tailscale-ip>:3100Then run the setup scripts:
# MCP server — gives Claude notify_user/ask_user tools
./scripts/setup-mcp.sh
# Notification hook — alerts you on tool approval prompts
./scripts/setup-hooks.shBoth scripts configure Claude Code at the user level (~/.claude.json and ~/.claude/settings.json), so notifications work across all sessions and projects. Restart any running Claude Code sessions to pick up the changes.
| Component | Technology |
|---|---|
| Backend | Node.js, Express, TypeScript |
| Frontend | React 19, Vite, TypeScript |
| Terminal | xterm.js, node-pty |
| Session backend | tmux |
| Database | SQLite (better-sqlite3) |
| Auth | JWT via jose |
MIT