Overview

Codemux is an agent development environment (ADE). It combines native multi-provider chat, terminal multiplexing, an embedded browser, git integration, and multi-agent orchestration in one desktop application. Agent Chat is the default interface; the classic terminal-first UI remains available in Settings.

What It Does

• Prompt-first workspaces — Describe what you want to do, and Codemux creates a workspace with an auto-generated branch, selected agent, and your task as the initial prompt. Image paste works directly in the prompt input.
• Project management — Open existing repos or create new ones (empty or clone), straight from the empty Home with your recent projects one click away. In the classic terminal-first flow, a first-project wizard detects your package manager and sets up dependencies automatically.
• Terminal multiplexer — Split panes, multiple tabs, WebGL rendering with Unicode 11 widths. Run AI coding agents (Claude Code, Codex, OpenCode, Gemini, Antigravity, Copilot, Cursor Agent, Amp, Grok, Droid, Mastracode, Pi) side by side. Panes persist across workspace switches so alt-screen TUIs keep their state.
• Agent status indicators — Real-time status dots show what each agent is doing: working (amber), needs input (red), ready for review (green), or monitoring something in the background (cyan). Visible in the sidebar and tab bar.
• Agent launcher — Use the title-bar + menu for Agent Chat, CLI-agent presets, terminals, and browser panes; pin favorite presets for one-click launch. A Run button with Ctrl+Shift+G executes your project's dev command instantly, and Set Run suggests commands detected from package.json, Makefiles, justfiles, and more.
• Embedded browser — Hybrid browser with three input tiers: CSS selector, CDP coordinate-based, and OS-level (ydotool / Win32 SendInput) for stealth interaction. Agents control it programmatically. Viewport presets resize the actual viewport via CDP for responsive testing.
• Git workflow — Stage, commit, push, pull, view diffs — all from the right panel. Works with GitHub and GitLab, self-hosted included.
• Code review — Create a pull or merge request, read its checks, leave line-level notes, submit a verdict, answer and resolve review threads, and merge — without leaving Codemux. A dedicated page triages every request across the repos you have open, and failing checks or review comments can be handed straight to an agent.
• Theming — Five dark themes picked live from the command palette, a studio for building your own from two colors or a VS Code import, and per-surface typography.
• Usage accounting — Token and cost totals per provider, model, and day, read from your agent CLIs' own local history.
• AI tools — Generate commit messages with AI from your staged diff.
• Setup scripts — Configure per-project setup/teardown commands that run automatically when workspaces open or close. Docker Compose support built in.
• Auth and sync — Sign in with GitHub OAuth or email/password. Personal settings and user-authored skills sync server-side across devices.
• Web remote access — Use the full Codemux UI from another browser. Pair with a running desktop, or run codemux serve on a displayless machine to boot the same backend without a GUI.
• Agent Chat — Default native conversational pane for Claude, Codex, Cursor, Grok, OpenCode, (experimentally) Hermes, and custom ACP agents, with streaming, mode pills, attachments, provider-authored task plans, live context-window usage, inline HTML visualizations, and a unified multi-provider model picker. Codex can run on a connected ChatGPT plan without an API key.
• Delegation and devices — A chat can hand a task to another installed agent or to a registered SSH checkout and get the result back in the thread, and a whole thread can run on another machine you've added as a device.
• Workflow orchestration — Dynamic workflows coordinate tasks across providers, keeping file changes isolated until you apply them, and a Claude Workflow run renders as an in-thread approval/progress card with a per-phase agent drill-in. See Workflow Orchestration.
• MCP control — Outside assistants can drive Codemux over its MCP server, including creating and controlling native chat threads once you approve them.
• Cross-platform — Native Linux (AppImage / .deb / .rpm / AUR) and Windows (NSIS installer) binaries. In-app auto-update covers the AppImage and the Windows NSIS build; .deb / .rpm / AUR update through their package manager.

Architecture

• Frontend: React 19 + Tailwind v4 + shadcn/ui, state via Zustand
• Backend: Rust (Tauri 2), async via Tokio
• Terminal: xterm.js with WebGL rendering (DOM fallback), Unicode 11 widths, kitty keyboard protocol, PTY via upstream portable-pty 0.8.1 (the Windows fork was reverted; the console is hidden by an AllocConsole + SW_HIDE on the parent process)
• Browser: Pure-Rust agent-browser via direct CDP. Hybrid 3-tier input (selector → CDP coordinates → OS-level ydotool / Win32 SendInput), screenshot-based rendering with reactive stream-url reconnect
• Git: Direct git CLI calls from Rust; GitHub via the gh CLI and GitLab via glab
• Agent Chat: AgentProvider trait with one adapter per provider — Claude (Claude Agent SDK via Bun-compiled sidecar), Codex (codex app-server JSON-RPC), OpenCode (Rust-direct HTTP against a managed opencode serve child), plus Cursor, Grok, Hermes, and custom ACP agents

Key Differences

Chat-first, terminal-complete: Native Agent Chat is the default, while real terminal panes remain first-class: they resize, split, and swap fluidly, use WebGL where a startup probe confirms it is beneficial, and fall back to DOM automatically.

Agent-native: Built for AI coding agents. Builtin presets launch any major coding-agent CLI — Claude Code, Codex, OpenCode, Gemini, Antigravity, Copilot, Cursor Agent, Amp, Grok, Droid, Mastracode, Pi — with one click, each in its own skip-permissions mode. Browser automation lets agents test their own output. Delegation and workflow orchestration coordinate multiple agents — even from different vendors — on a single task.

Workspace isolation: Each workspace can use git worktrees so agents work on separate branches without interfering with each other.