supermux

AFK. Still shipping.

supermux runs Claude Code, Codex, Cursor & OpenCode 24/7 on a box you own —
every session, on every screen you carry.

Website · Install · Setup guide

Agents only work while you watch

Coding agents only make progress while you sit in front of them. Close the laptop and the session dies; a finished diff waits hours for review; one question stalls a task until evening.

supermux gives your agents a box of their own — a VPS, a mini PC, the spare laptop in a drawer — where they run around the clock. The web app (and Telegram) is your window into all of them: spawn a session from your phone, answer a worker's question from the train, review the diff after dinner, merge when it's right.

No vendor cloud, no account. Your agent subscriptions, your code, your box.

curl -fsSL https://supermux.dev/install.sh | sh

_{Windows (PowerShell): irm https://supermux.dev/install.ps1 | iex — more options under Install.}

What changes for you

• Leave whenever. Sessions live on the box, not your laptop — the lid closing changes nothing.
• Never babysit. A push finds you when a worker finishes or asks; one tap opens that exact session.
• Review from your phone. Each session works in its own git worktree — read the diff, comment inline, the agent fixes, and main moves only on your word.
• Autonomous, not reckless. No permission prompts; isolated worktrees, a dedicated box, and nothing merges unreviewed.
• It learns your stack. Shared memory across sessions plus a soul.md identity you write — knowledge compounds instead of resetting.

Your window into every session

• A chat per session. An installable PWA with one thread per session — voice input with a live waveform, photos, camera, and push that stays quiet while you're already looking. Telegram works too, when chat is all you need.
• Code review, pocket-sized. Read the full diff grouped by repo, drop inline comments, the agent addresses them — main fast-forwards only when you say so.
• A full workspace in the browser. File tree, editor with search, and a real terminal (xterm.js PTY) in the session's directory. Rate limits and spend across every agent in one dashboard.

Four agents, one interface

Claude Code, Codex, Cursor, and OpenCode drive identically — same chat, same workspace, same orchestration. Switch models live mid-session, dial thinking effort up for the hard problems, and mix agents freely across sessions. If your subscription changes, your workflow doesn't.

How it works

  phone · tablet · laptop               a box you own
 ┌─────────────────┐        ┌──────────────────────────────────────┐
 │  web app (PWA)  │        │  supermux broker                     │
 │  Telegram       │ ◀────▶ │  routing · push · files · proxy      │
 └─────────────────┘        │      │                               │
                            │   tmux sessions                      │
                            │ ┌──────┐ ┌─────┐ ┌──────┐ ┌────────┐ │
                            │ │claude│ │codex│ │cursor│ │opencode│ │
                            │ └──────┘ └─────┘ └──────┘ └────────┘ │
                            └──────────────────────────────────────┘

• One daemon, your keys. The broker owns your channel credentials and supervises every session in tmux. Control commands (/spawn, /switch, …) run in the broker itself, so routing burns no agent tokens.
• Sessions that survive. Restart the broker or reboot the box — every session resumes with full history. Assistant sessions persist and orchestrate; workers are spawned per task.
• Yours, end to end. No vendor cloud, no account. Reach your box over LAN, your own reverse proxy, a tunnel, or a VPN — your call.

Install

One command, and it walks you through the rest.

Mac / Linux

curl -fsSL https://supermux.dev/install.sh | sh

Windows (PowerShell)

irm https://supermux.dev/install.ps1 | iex

The installer offers Docker or a native install, configures connectivity, and pairs your first device. Prefer to do it by hand? Either of these also works:

• Docker — git clone https://github.com/UstaLabs/supermux.git && cd supermux && docker compose up -d, then open http://localhost:8787. State persists in named volumes; drop the repos you want worked on into ./workspace. Uncomment image: ghcr.io/ustalabs/supermux:latest to skip the local build.
• Guided — paste SETUP.md into Claude Code, Claude.ai, or ChatGPT; it asks where to run supermux (this computer, a mini PC / VPS over SSH, or a throwaway container) and walks you through it.

On first open, a setup wizard covers agents (paste a CLAUDE_CODE_OAUTH_TOKEN from claude setup-token, an Anthropic/OpenAI API key, or a device-flow login), connectivity (optional Telegram token, a public URL with ready-to-paste Caddy / nginx / Cloudflare Tunnel snippets), and your first session. The first browser to open a fresh instance is paired automatically; on a headless box, bun run pair <name> does the same.

Configuration

Set these in ~/.mux/state/.env (native) or .env (Docker):

Requirements

• macOS (Apple Silicon) or Linux natively — or the Docker image on any host. Windows runs it through WSL2 (the PowerShell installer sets that up). An always-on box — VPS, mini PC, or home server — is where it shines; your daily machine works too.
• tmux on PATH, plus at least one agent CLI logged in: claude, codex, cursor-agent, or opencode.
• Native installs also need bun and git. The broker runs fine web-only (no Telegram token) — just set MUX_WEB_PORT + MUX_WEB_PUBLIC_URL.

Built with supermux

Every feature above was designed, implemented, and reviewed by agent sessions running inside supermux — the research, the specs, the code, the review. This README included, steered from a tablet. We're the first users of every build.

License

MIT — free and open source.