wmux is LSP-for-terminals — a neutral substrate that lets external tools build workflow intelligence on top of any terminal session. Native Windows terminal multiplexer with split panes, MCP bridge, and browser automation — purpose-built for running Claude Code, Codex CLI, and Gemini CLI side by side. No WSL required.
Keywords: Windows tmux, tmux for Windows, terminal multiplexer Windows, AI agent terminal, Claude Code Windows, Codex CLI, Gemini CLI, MCP server, Chrome DevTools Protocol, split terminal, multi-agent terminal, browser automation, ConPTY, xterm.js, Electron terminal.
Windows has no native tmux. Without WSL, there was no clean way to run multiple AI coding agents side by side.
wmux fixes this. A native Windows terminal multiplexer + browser automation + MCP server, purpose-built for AI coding agents like Claude Code, Codex CLI, and Gemini CLI. Your AI agent reads the terminal, controls the browser, and works autonomously — all in one window.
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.
Winget (recommended — no security warning):
winget install openwong2kim.wmuxChocolatey:
choco install wmuxInstaller: Download wmux Setup.exe
Seeing a "Windows protected your PC" warning? The installer is not yet Authenticode-signed (free code signing via SignPath is being set up — see Code signing below), so Windows SmartScreen flags it as from an unknown publisher. It's safe to proceed — click More info → Run anyway. Installing via winget or Chocolatey above avoids this prompt entirely, since those package managers run in a trusted context.
One-liner (PowerShell): downloads the prebuilt Setup.exe from the latest release and verifies its SHA-256 before launching it (no Node/Python/build tools needed).
irm https://raw.githubusercontent.com/openwong2kim/wmux/main/install.ps1 | iexWant to build from source instead (needs Node 18+, Python 3, and VS C++ Build Tools — auto-installed)?
$env:WMUX_FROM_SOURCE=1; irm https://raw.githubusercontent.com/openwong2kim/wmux/main/install.ps1 | iexCode signing — Free Authenticode code signing for Windows builds is provided by SignPath.io, certificate by the SignPath Foundation.
Tell Claude Code "search Google for this" and it actually does it. wmux's built-in browser connects via Chrome DevTools Protocol (CDP). Click, type, screenshot, execute JS — all done by the AI directly. Works perfectly with React controlled inputs and CJK text (Korean, Japanese, Chinese).
You: "Search for wmux on Google"
Claude: browser_open → browser_snapshot → browser_fill(ref=13, "wmux") → browser_press_key("Enter")
→ Actually searches Google. Done.
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 pseudo-terminal. xterm.js + WebGL hardware-accelerated rendering. 999K lines of scrollback. Terminal content persists even after app restart.
Tmux-style prefix mode — Ctrl+B then a single action key (split, focus, new workspace…) for muscle-memory terminal navigation. 13 actions, fully rebindable.
Floating pane — Ctrl+` for a Quake-style dropdown terminal that lives outside your main layout. Stays alive across toggles.
Scroll bookmarks — Ctrl+M marks the current scroll position, Ctrl+Up/Down jumps between marks. Indicators render on the gutter.
Smart right-click — Windows Terminal style. Selection → instant copy. Empty area → instant paste. Link → small Open / Copy Link menu. Zero modal interrupts.
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.
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 |
| Read commands semantically (OSC 133) | terminal_read_events |
| Send command | terminal_send |
| Manage workspaces | workspace_list / surface_list / pane_list |
| Agent-to-agent messaging | a2a_send / a2a_broadcast / a2a_whoami |
| Delegate tasks across agents | a2a_task_send / a2a_task_query / a2a_task_cancel |
| Company mode coordination | company_a2a_send / company_a2a_inbox / company_a2a_status |
Multi-agent: Every browser tool accepts surfaceId — each Claude Code session controls its own browser independently. A2A (agent-to-agent) tools route messages between sibling agents in the same wmux instance.
Programmatic orchestration: If you want to drive multiple wmux panes from your own script (TypeScript / Node), the MCP surface is wrapped by @wmux/orchestrator — a small client-side SDK that gives you sendAndWait, sendCommandAndWait (OSC 133 exit codes), atomic pane claim, fan-out, and agent.lifecycle event subscription. Composes the primitives above, ships independently from wmux.
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.
- 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 — cookie encryption on; Node CLI inspect args and
NODE_OPTIONSenv disabled; app loads only from asar. (RunAsNodestays enabled — the background daemon is spawned as a detached Node process fromwmux.exeviaELECTRON_RUN_AS_NODE=1.)
- xterm.js + WebGL GPU-accelerated rendering
- ConPTY native Windows pseudo-terminal
- Unicode 11 width tables — correct CJK / emoji rendering for cursor-positioning TUIs (Claude Code, vim)
- Split panes —
Ctrl+Dhorizontal,Ctrl+Shift+Dvertical - Tabs — multiple surfaces per pane
- Floating pane — Quake-style dropdown terminal, dedicated PTY,
Ctrl+` - Smart right-click — selection → instant copy, empty area → instant paste, link → Open / Copy Link menu
- Scroll bookmarks —
Ctrl+Mmark,Ctrl+Up/Downjump, gutter indicators - Vi copy mode —
Ctrl+Shift+X - Search with regex toggle —
Ctrl+F - 999K line scrollback with disk persistence
- Shell integration (OSC 133) — semantic prompt / command boundaries for
terminal_read_events. Auto-injected for pwsh / bash. Constrained Language Mode safe (v2.7.1).
- Tmux-style prefix mode —
Ctrl+Bthen action key, 13 default actions (splits, focus, workspaces, palette, floating pane, …) - Customizable bindings + custom keymaps in Settings
- Reset-to-defaults available
- Sidebar with drag-and-drop reordering
Ctrl+1~9quick switch- Multiview —
Ctrl+clickto view multiple workspaces side by side - Layout templates — save current pane layout, restore via Command Palette ("recent" category)
- Full session persistence — layout, tabs, cwd, scrollback all restored
- One-click reset in Settings
- 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
- Output throughput-based activity detection
- Taskbar flash + Windows toast notifications
- Process exit alerts
- Notification panel —
Ctrl+I - Web Audio sound effects
Claude Code, Cursor, Aider, Codex CLI, Gemini CLI, OpenCode, GitHub Copilot CLI
- Detects agent start → activates monitoring
- Critical action warnings
- Background session management (survives app restart)
- Scrollback buffer dump and auto-recovery
- Windows startup registration (survives reboot)
- Dead session TTL reaping (24h default)
- Automatic update checks via GitHub Releases
- Toggle on/off in Settings > General
- Manual check available in Settings
Catppuccin Mocha, Monochrome, Sandstone
English, Korean, Japanese, Chinese
| 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 then <key> |
Tmux-style prefix mode (13 actions) |
Ctrl+Shift+B |
Toggle sidebar |
Ctrl+` |
Toggle floating pane (Quake-style) |
Ctrl+K |
Command palette |
Ctrl+I |
Notifications |
Ctrl+, |
Settings |
Ctrl+F |
Search terminal (with regex toggle) |
Ctrl+M |
Add scrollback bookmark |
Ctrl+Up / Ctrl+Down |
Jump prev / next bookmark |
Ctrl+Shift+X |
Vi copy mode |
| Right-click | Smart copy (selection) / paste (empty) / link menu |
F12 |
Browser DevTools |
git clone https://github.com/openwong2kim/wmux.git
cd wmux
npm install
npm start # Dev mode
npm run make # Build installer- 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 only when building from source (-FromSource / WMUX_FROM_SOURCE=1). The default one-liner downloads the prebuilt Setup.exe and needs none of these.
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
Is wmux a tmux port for Windows? No. wmux is a native Windows terminal multiplexer built on ConPTY and Electron, offering tmux-style split panes, prefix keys, and session persistence — without requiring WSL or Cygwin.
Does wmux work with Claude Code, Codex CLI, and Gemini CLI? Yes. wmux automatically detects these AI coding agents and registers an MCP server so Claude Code can drive the built-in browser and read terminal output.
Can I run multiple AI agents at the same time? Yes. Each pane runs an independent PTY. Agents can communicate via the A2A (agent-to-agent) MCP tools for multi-agent workflows.
Does it require WSL? No. wmux is fully native Windows (ConPTY + Electron). No WSL, Cygwin, or MSYS2 needed.
Why does Windows show a "Windows protected your PC" warning when I run the installer?
The installer isn't code-signed yet, so Windows SmartScreen flags it as coming from an unknown publisher. It's safe to proceed — click More info → Run anyway. To avoid the prompt entirely, install via winget install openwong2kim.wmux or choco install wmux; those package managers run in a trusted context, so SmartScreen never appears.
- xterm.js — Terminal rendering
- node-pty — Pseudo-terminal
- Electron — Desktop framework
- Playwright — Browser automation engine
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.
MIT