Wmux: workspace multiplexer for running AI coding agents in parallel

Wmux is a workspace multiplexer for AI coding agents, released on Hacker News and available for Windows, macOS, and Linux. It lets developers run multiple coding agents—Claude Code, OpenAI Codex, Google Gemini, Aider, OpenCode, and GitHub Copilot CLI—in parallel panes within a single window, with full session recovery across crashes and OS reboots.

The core architecture is a daemon that owns every PTY (ConPTY on Windows, forkpty on macOS), rendering through a WebGL terminal (xterm.js) with Unicode 11 support, 999K-line scrollback, and tmux-style keyboard shortcuts. Panes can be split horizontally or vertically, grouped into tabs, and arranged across multiple workspaces; a "Fleet View" cockpit (Ctrl+Shift+A) shows all agents across all workspaces in an always-on side panel, with blocked agents floated to the top. Declared panes in a wmux.json file are supervised like an init system—auto-restarted on crash and reboot, with restart policies and backoff—and any pane declared in that file is resumed at app launch.

For Git integration, a dock tab shows the worktrees of the repo behind the active pane, plus its pull requests and comments (GitHub via gh, GitLab via glab, including self-hosted). A read-only workspace diff—showing every staged, unstaged, and untracked change against HEAD—is one palette command away. From any diff hunk, users can ask the orchestrator a question with the repo, file, and code fenced into the message. Task fan-out is built in: users point the orchestrator at an objective and can spawn up to 8 WorkTask missions, each with a dedicated git worktree on a fresh wtask/* branch, its own task workspace, a private mission channel, and a file-backed initial prompt. Worktrees are never force-deleted—if a task fails, dirty output is preserved. Users harvest results through a diff surface with per-hunk checkboxes; adoption is a single all-or-nothing git apply. When done, a task can be closed (the worktree is removed only after a clean check) or opened as a pull request in one click.

Multi-agent orchestration is supported through agent-to-agent (A2A) messaging and channels. Agents can message and delegate tasks by pane, gated by a per-pane execute approval gate—a remote agent cannot spawn code in the user's workspace without explicit OK. Channels are Slack-style rooms for a workspace's agents: users can create, join, invite, post, read, and archive; each message carries a server-verified sender shown as a pane identity chip plus a per-workspace color badge. Unread counts and @-mention badges survive reboot. A headless wmux channel CLI (unread / read / post / ack / join / list) lets agents and operators interact via the command line.

Agents control a built-in Chrome browser over the Chrome DevTools Protocol (CDP). The browser panel includes a nav bar, DevTools, back-forward buttons, a hover-highlight element inspector for copying LLM context, and full automation: click, fill, type, screenshot, JavaScript eval, and key press. All 84 MCP tools (browser_open, terminal_read, pane_list, a2a_task_send, channel_post, and others) register themselves at launch with zero configuration when Claude Code is running.

Notifications are native: desktop toasts and taskbar flash (Windows) or Dock and menu-bar tray (macOS) on agent completion. Activity detection is based on output throughput (not pattern matching), so it works with any agent. A notification panel (Ctrl+I) and wmux web (a PWA-installable read-only phone browser view of live panes) provide additional visibility. The tool flags destructive actions (rm -rf, git push --force, DROP TABLE) for approval.

Installation is streamlined. Windows users install via winget (winget install openwong2kim.wmux), choco, or download Setup.exe for offline install; the installer currently shows an unknown-publisher warning in SmartScreen but winget and choco skip the prompt. macOS users download a Developer ID signed, notarized, and stapled .dmg for Apple Silicon and drag wmux to Applications; the CLI installs itself on first launch. Both platforms auto-update in place: wmux checks for a new release every 30 minutes and verifies downloads against a published SHA-256 before installing (Windows x64 and macOS arm64 supported). Linux has experimental AppImage, .deb, and .rpm builds on the releases page.

UI customization includes 10 themes (Amber by default, plus Catppuccin, Nightowl, Monochrome, Void, Hinomaru, Taegeuk, Stars & Stripes, Red Dynasty, and Custom) and 10 terminal palettes, with light options included. Localization is scaffolded for 23 locales; English and Korean are complete, and Japanese and Chinese are in progress—translations are welcome. Security features include token-authed IPC, SSRF guards, PTY input sanitization, a randomized CDP port, and Electron Fuses.

Wmux addresses a growing coordination problem in AI-assisted development: running multiple coding agents in parallel without losing context or requiring manual orchestration. The tool's core innovation is treating the workspace—not just the chat—as a first-class object, bundling the terminal PTYs, git state, browser automation, and inter-agent messaging into a single daemon-backed session. This persistence model, borrowed from tmux, allows teams to treat agent work as durable across system events, a critical requirement for long-running orchestration tasks.

The feature set targets two distinct workflows. The first is prompt fan-out: splitting a single objective into up to 8 isolated worktrees, each handled by a different agent, then reviewing and merging results through a diff interface with per-hunk adoption (all-or-nothing git apply). This avoids the manual conflict-resolution tax of traditional PR workflows. The second is multi-agent coordination: agents can message each other, delegate tasks, and read shared channels (modeled on Slack), all gated by approval checkpoints so no agent can run code without explicit permission. Both workflows are hedged against agent failure and restart—missions record decisions in channels, worktrees are never force-deleted on failure, and dirty output is preserved.

No comments yet. Be the first to share your thoughts!