duxweb/codux

Codux – High‑performance AI‑coding terminal (archived)

What it is

• A native desktop (macOS, Windows, Linux) and mobile app that acts as a control plane for developers who use AI‑powered coding assistants (e.g., OpenAI Codex, Claude Code, Kimi, etc.).
• It bundles a terminal, project‑aware workspace, token‑usage analytics, credential‑isolated SSH/DB helpers, and encrypted peer‑to‑peer links so you can keep long‑running AI‑agent sessions alive across devices and headless hosts.

Core ideas

Problem	Codux’s answer
Each AI CLI keeps its own state → chaos across worktrees, terminals, tokens	One unified workspace that tracks every AI CLI, Git worktree, and terminal per project
Long agent runs are hard to resume	Live status, local history, and session restore; you can detach and re‑attach from phone or another computer
Token spend is opaque	Real‑time token usage broken down by tool, model, project, and day
Credentials end up in prompts	codux-ssh and codux-db wrappers inject passwords/keys inside a protected helper process – the model never sees them
Work lives on a remote machine	A headless host (codux-agent) runs on any server/Mac/Linux box; desktop or phone connects via end‑to‑end‑encrypted P2P/relay links

Key features

• Multi‑tool support – wrappers for Codex, Claude Code, Oh My Pi, OpenCode, MiMo, Kiro, CodeWhale, Agy, Kimi, etc.
• Worktree‑first model – each Git worktree gets its own terminals, Git state, and AI sessions.
• Token analytics – dashboards show usage per model/tool/project.
• Credential isolation – codux-ssh / codux-db commands let agents access servers/databases without exposing passwords in prompts or shell history.
• Device‑agnostic pairing – desktop, phone, and any headless host become peers via the iroh networking library; connections are encrypted and can fall back to relays.
• Web Tunnel Browser – a built‑in Chromium view that proxies web traffic through the host, letting you browse a Vite dev server on the remote machine as if it were local.
• Terminal pet – a whimsical pixel pet that levels up based on token consumption (purely cosmetic).
• Local‑first – all data (projects, memory, credentials, token stats) stays on your machines; there is no cloud service or account.

Architecture (high level)

• Desktop app – written in Rust + GPUI, provides the UI, terminal emulator, and the wrappers that inject environment directives into supported AI CLIs.
• Mobile app – a Flutter front‑end that connects to the same peer network, showing the same workspace.
• Headless host (codux-agent) – a lightweight Rust binary that runs on any OS, exposing terminals, Git, and AI sessions over the encrypted link.
• Networking – powered by the iroh library for end‑to‑end encrypted P2P connections with optional relay fallback.
• Credential helpers – small subprocesses (codux-ssh, codux-db) that read stored profiles and perform the actual connection, keeping secrets out of the AI model’s context.

Installation

• Desktop – macOS via Homebrew Cask (brew install --cask duxweb/tap/codux) or direct download of a .dmg / Windows .exe installer.
• WSL on Windows – enable WSL, install a Linux distribution, then install Codux Runtime from the app’s Settings.
• Mobile – Android APK from the codux-flutter releases; iOS app is paid (covers Apple developer fees) but source is open‑source.
• Headless host – one‑liner script curl …/install.sh | sh (auto‑detects OS/arch) or manual binary download.

License & status

• Licensed under GPL‑3.0.
• The repository is archived; active development has moved to the Sailry project (sailry-harness). The code remains available for reference.

Who might use it

• Developers who run AI‑assisted coding tools heavily and need a stable, project‑scoped environment.
• Teams that want to keep long‑running AI agents on a server while still being able to monitor and interact from any device.
• Anyone concerned about leaking credentials into LLM prompts.

---

All information above is taken directly from the README; no additional features are inferred.