# luxcode.org

> lux: a coding agent in a Luxor box, one command away, with auto-resume
> and your choice of inference provider.

lux is a single bash script. An "agent" is a name + harness + inference
provider (+ model); each agent has its own Luxor sandbox. Harnesses: terminal
agents claude, codex,
opencode, copilot, gemini, qwen, crush, goose, kimi, pi, omp, droid, adal,
hermes, openclaw, plus Luxor's pre-built images (tenki.cloud/docs/sandbox/
base-image), whose box is created from that image: grok (Grok Build, XAI_API_KEY
or device login), amp (AMP_API_KEY or amp login), droid and pi (pinned builds),
prime-agent (PRIME_API_KEY or login), and the toolkits tinyfish, anchor, hotdata
(a shell with the key set; `run --` passes raw CLI arguments) — all in tmux,
attached over ssh — and web UIs opencode-web,
open-webui, hermes-web, openclaw-web, durable-pi (served from the box, tunnelled
to localhost; `lux <agent> url` makes a public Luxor preview link). Creating an
agent asks terminal-or-web first, then the harness. Each agent has its own Luxor sandbox
named after it, which pauses by itself 2h after each create or resume (Luxor's
cap on max-duration; the CLI has no idle timeout); the next lux command
resumes it. COLINCODE_STICKY=1 creates boxes that never pause by themselves. lux creates or resumes the sandbox, starts the agent
inside tmux with the provider wired in, and attaches you to it.

## Install

    curl -fsSL https://luxcode.org/install.sh | sh

Requires the Luxor CLI (https://luxor.ai) and jq. The raw script is served at
https://luxcode.org/lux.

## Commands

    lux                      pick an agent (or create one) and jump into it
    lux new [cli|web|harness] create: terminal or web -> harness -> provider -> model -> size -> volume -> name
    lux <harness> [action]   a harness (list above)
    lux <agent> [action]     a specific agent

    up (default)   create or resume the box, start the agent, attach
    ssh            same as up
    shell          plain shell in the box
    run -- <p>     one-shot headless prompt
    status         agent, provider, box state
    provider       switch inference provider (and model)
    model [id]     switch model
    size [name]    box size: small 2vCPU/4GiB, medium 4/8 (default), large 8/16,
                   xlarge 12/32, xxlarge 16/64; applies when the box is (re)created
    volume [GiB]   persistent volume at ~/work (10/25/50 GiB, or none; the workspace caps the size); the
    url [off]      web UIs only: a public Luxor preview link to the UI (no password of its own), or remove it
    image [build [--full|--public]|forget]  the golden image (Luxor base + tmux) new boxes start from: found or built on first use
                   agent starts there; survives pause, rm and resize. Volumes keep
                   their size for life; the root disk is wiped by rm.
    auth           sign in (claude: `claude auth login` in the box) / store key
    down | pause   pause the box (auto-resumed on next command)
    rm | stop | kill   terminate the box and forget the agent (--purge deletes the volume)
    lux ls       list agents
    lux pause --all   pause every agent's box
    lux rm --all      terminate every box and forget every agent
                        (stop/kill/terminate --all do the same)

## Providers

Harness defaults: each harness's own sign-in or in-app setup (Anthropic for
claude, the only one with Remote Control; OpenAI for codex; GitHub for copilot;
Google for gemini; Qwen; Moonshot for kimi; Factory for droid; the rest are
configured inside). Gemini CLI and AdaL take only their own. Others: Nebius
Token Factory, Boundless, Baseten, Friendli, OpenRouter (every other harness;
written into the harness's own config, which also puts that harness in its
no-approval mode; Claude Code loses Remote Control on
any third-party base URL). Nebius has no Anthropic Messages API, so for Claude
Code lux runs a stdlib-only Python translation proxy inside the box (tmux
session claude-proxy, 127.0.0.1:8787, port via COLINCODE_PROXY_PORT) that serves
/v1/messages with streaming and tool calls in front of /chat/completions. Default
models are each provider's most advanced: Kimi K3 (Nebius, Baseten), Claude Fable
5.1 (Boundless, OpenRouter), GLM 5.3 (Friendli). Keys: NEBIUS_API_KEY,
BOUNDLESS_API_KEY, BASETEN_API_KEY, FRIENDLI_TOKEN, OPENROUTER_API_KEY in
~/.config/luxcode/env.

## Configuration

COLINCODE_HARNESS (claude), COLINCODE_SIZE (medium), COLINCODE_VOLUME_GB (10),
COLINCODE_CPU (4), COLINCODE_MEMORY_MB (8192),
COLINCODE_DISK_GB (20), COLINCODE_IMAGE (none), COLINCODE_RESUME_TIMEOUT (3m),
COLINCODE_MAX_DURATION (2h, the cap), COLINCODE_PAUSE_RETENTION (720h), COLINCODE_STICKY (0).

## Links

- Source: https://github.com/opencolin/luxcode
- Sandboxes: https://luxor.ai
- Related: https://nemocode.org (inference relay for coding agents)