wmux — AI Agent Terminal for Windows

Run Claude Code + Codex + Gemini CLI side by side. Split terminals, browser automation, MCP integration — the only proper way to use AI agents on Windows.

---

Still using one terminal for your AI coding agents on Windows?

macOS has cmux — a tmux-based terminal multiplexer for AI agents.

Windows has no tmux. Without WSL, there was no way.

wmux fixes this. Native Windows terminal multiplexer + browser automation + MCP server. Your AI agent reads the terminal, controls the browser, and works autonomously.

Claude Code writes the backend on the left
Codex builds the frontend on the right
Gemini CLI runs tests at the bottom
— all on one screen, simultaneously.

---

Install in 30 seconds

Installer:

Download wmux Setup.exe

One-liner (PowerShell):

irm https://raw.githubusercontent.com/openwong2kim/wmux/main/install.ps1 | iex

---

Why wmux?

1. Your AI agent controls the browser — for real

Tell Claude Code "search Google for this" and it actually does it.

wmux's built-in browser connects via Chrome DevTools Protocol. Click, type, screenshot, execute JS — all done by the AI directly. Works perfectly with React controlled inputs and CJK text.

You: "Search for wmux on Google"
Claude: browser_open → browser_snapshot → browser_fill(ref=13, "wmux") → browser_press_key("Enter")
→ Actually searches Google. Done.

2. Multiple terminals in one window

Ctrl+D to split, Ctrl+N for new workspace. Place multiple terminals and browsers in each workspace. Ctrl+click for multiview — see multiple workspaces at once.

ConPTY-based native Windows terminal. xterm.js + WebGL hardware-accelerated rendering. 999K lines of scrollback. Terminal content persists even after restart.

3. No more asking "is it done yet?"

wmux tells you when your AI agent finishes.

• Task complete → desktop notification + taskbar flash
• Abnormal exit → immediate warning
• git push --force, rm -rf, DROP TABLE → dangerous action detection

Not pattern matching — output throughput-based detection. Works with any agent.

4. Automatic Claude Code integration

Launch wmux and the MCP server registers automatically. Claude Code just works:

What Claude can do	MCP Tool
Open browser	browser_open
Navigate to URL	browser_navigate
Take screenshot	browser_screenshot
Read page structure	browser_snapshot
Click element	browser_click
Fill form	browser_fill / browser_type
Execute JS	browser_evaluate
Press key	browser_press_key
Read terminal	terminal_read
Send command	terminal_send
Manage workspaces	workspace_list / surface_list / pane_list

Multi-agent: Every browser tool accepts surfaceId — each Claude Code session controls its own browser independently.

5. Session persistence — like tmux

Terminal sessions survive app restarts. Close wmux and reopen — your sessions are still there, scrollback intact.

• App restart: Daemon keeps PTY processes alive in the background. Reconnects instantly.
• Reboot: Sessions are recovered with saved scrollback and working directory. wmux auto-starts on login.
• Auto-update: Checks for updates via GitHub Releases. Toggle on/off in Settings.

6. Security that actually matters

• Token authentication on all IPC pipes
• SSRF protection — blocks private IPs, file://, javascript: schemes
• PTY input sanitization — prevents command injection
• Randomized CDP port — no fixed debug port
• Memory pressure watchdog — reaps dead sessions at 750MB, blocks new ones at 1GB
• Electron Fuses — RunAsNode disabled, cookie encryption enabled

---

All Features

Terminal

• xterm.js + WebGL GPU-accelerated rendering
• ConPTY native Windows pseudo-terminal
• Split panes — Ctrl+D horizontal, Ctrl+Shift+D vertical
• Tabs — multiple surfaces per pane
• Vi copy mode — Ctrl+Shift+X
• Search — Ctrl+F
• 999K line scrollback with disk persistence

Workspaces

• Sidebar with drag-and-drop reordering
• Ctrl+1~9 quick switch
• Multiview — Ctrl+click to view multiple workspaces side by side
• Full session persistence — layout, tabs, cwd, scrollback all restored
• One-click reset in Settings

Browser + CDP Automation

• Built-in browser panel — Ctrl+Shift+L
• Navigation bar, DevTools, back/forward
• Element Inspector — hover to highlight, click to copy LLM-friendly context
• Full CDP automation: click, fill, type, screenshot, JS eval, key press

Notifications

• Output throughput-based activity detection
• Taskbar flash + Windows toast notifications
• Process exit alerts
• Notification panel — Ctrl+I
• Web Audio sound effects

Agent Detection

Claude Code, Cursor, Aider, Codex CLI, Gemini CLI, OpenCode, GitHub Copilot CLI

• Detects agent start → activates monitoring
• Critical action warnings

Daemon Process

• Background session management (survives app restart)
• Scrollback buffer dump and auto-recovery
• Windows startup registration (survives reboot)
• Dead session TTL reaping (24h default)

Auto-Update

• Automatic update checks via GitHub Releases
• Toggle on/off in Settings > General
• Manual check available in Settings

Themes

Catppuccin Mocha, Monochrome, Sandstone

i18n

English, Korean, Japanese, Chinese

---

Keyboard Shortcuts

Key	Action
Ctrl+D	Split right
Ctrl+Shift+D	Split down
Ctrl+T	New tab
Ctrl+W	Close tab
Ctrl+N	New workspace
Ctrl+1~9	Switch workspace
Ctrl+click	Add to multiview
Ctrl+Shift+G	Exit multiview
Ctrl+Shift+L	Open browser
Ctrl+B	Toggle sidebar
Ctrl+K	Command palette
Ctrl+I	Notifications
Ctrl+,	Settings
Ctrl+F	Search terminal
Ctrl+Shift+X	Vi copy mode
F12	Browser DevTools

---

Development

git clone https://github.com/openwong2kim/wmux.git
cd wmux
npm install
npm start           # Dev mode
npm run make        # Build installer

Requirements (dev only)

• Node.js 18+
• Python 3.x (for node-gyp)
• Visual Studio Build Tools with C++ workload

The install.ps1 script auto-installs Python and VS Build Tools if missing.

---

Architecture

Electron Main Process
├── PTYManager (node-pty / ConPTY)
├── PTYBridge (data forwarding + ActivityMonitor)
├── AgentDetector (gate-based agent status)
├── SessionManager (atomic save with .bak recovery)
├── ScrollbackPersistence (terminal buffer dump/load)
├── PipeServer (Named Pipe JSON-RPC + token auth)
├── McpRegistrar (auto-registers MCP in ~/.claude.json)
├── WebviewCdpManager (CDP proxy to <webview>)
├── DaemonClient (daemon mode connector)
├── AutoUpdater (GitHub Releases feed)
└── ToastManager (OS notifications + taskbar flash)

Renderer Process (React 19 + Zustand)
├── PaneContainer (recursive split layout)
├── Terminal (xterm.js + WebGL + scrollback restore)
├── BrowserPanel (webview + Inspector + CDP)
├── NotificationPanel
├── SettingsPanel (auto-update toggle, workspace reset)
└── Multiview grid

Daemon Process (standalone)
├── DaemonSessionManager (ConPTY lifecycle)
├── RingBuffer (circular scrollback buffer)
├── StateWriter (session suspend/resume)
├── ProcessMonitor (external process watchdog)
├── Watchdog (memory pressure escalation)
└── DaemonPipeServer (Named Pipe RPC + token auth)

MCP Server (stdio)
├── PlaywrightEngine (CDP connection, fast-fail)
├── CDP RPC fallback (screenshot, evaluate, type, click)
└── Claude Code <-> wmux Named Pipe RPC bridge

---

Acknowledgments

• cmux — The macOS AI agent terminal that inspired wmux
• xterm.js — Terminal rendering
• node-pty — Pseudo-terminal
• Electron — Desktop framework
• Playwright — Browser automation engine

---

Note on AI Agents

wmux detects AI coding agents for status display purposes only. It does not call any AI APIs, capture agent outputs, or automate agent interactions. Users are responsible for complying with their AI provider's Terms of Service.

License

MIT