Lux Code – terminal AI coding agent
Lux Code is an AI coding agent for the terminal. Tell it what you need and it works in your
project: it reads the code, edits files, runs commands and tests in the LuxOS shell lsh and
continues from the results. It uses the same assistant and the same account as LuxAI,
but works in the terminal and is aimed at programming. Since LuxOS 1.0.3 Lux Code (version 0.6) is
built on the open-source Oh My Pi (MIT) with a LuxOS layer: the LuxAI account and
models, lsh as the shell, the LuxOS manual and LuxOS rules.
Install and run
lux install lux-code # from the [luxos-apps] repository (needs bun from Arch, installed with it)
cd my-project
lux-code # interactive; lux-code -p "TASK" for a single task
| Item | Value |
|---|---|
| Package | lux-code 0.6.5 (replaces lux-code-next and the original Lux Code 0.1) |
| Command | lux-code (also a menu entry Lux Code); lux-code-next still works |
| Settings | ~/.config/lux-code/agent/config.yml (also editable with /settings) |
| Project instructions | AGENTS.md directly in the working directory only (not parent folders) |
| Docs on disk | /usr/share/doc/lux-code/README.md |
Settings of the earlier test version in ~/.config/lux-code-next are moved to
~/.config/lux-code on the first start. The original Lux Code 0.1 kept its data in
~/.local/share/lux-code; it is left untouched.
Sign-in
On LuxOS, Lux Code uses the shared LuxAI account (see LuxAI account); if
you are not signed in, run /luxai login or luxos-account login. Requests count toward the
account's monthly credits. Without the account Lux Code can use an API key (LUX_AI_API).
Terminal interface
| Command | Purpose |
|---|---|
/help |
all commands |
/model flash, /model lite |
choose LuxAI Flash or Lite |
/usage |
LuxAI credits this month (bar, credits left, renewal date) |
/luxai, /luxai login |
account status and personalization; sign in |
/cd DIR |
change the working directory (/cd = home, /cd - = previous, Tab completes) |
/settings |
settings, e.g. turning off LuxAI personalization for Lux Code |
/hub or Alt+A |
what the subagents are doing |
/notifications on\|off |
desktop notifications |
/mouse off |
classic inline view with the terminal's own scrollback |
- The model's thinking is hidden behind one line; Ctrl+T shows it.
- The mouse works like in other coding agents: drag to select and copy, double click selects a word, triple click a line, the wheel and PgUp/PgDn scroll the conversation.
- Subagents run in the background (at most 3 at a time) while the main agent keeps working.
- Long sessions are summarized automatically before they reach the LuxAI request limits.
Tools
The model gets a small core set (files, edits, search, shell commands in lsh, web) and loads the
rest when needed with tool_search: desktop and screenshots, clipboard, Android via adb, Firefox
debugging, system information, language servers and the offline LuxOS and LSL manual from
/usr/share/luxos/docs. Commands that LuxOS does differently are redirected: pacman, yay and
similar are refused with the matching lux command, and passwords are never piped to sudo.
Lux Code is not a sandbox: an approved shell command runs with your normal rights. For untrusted code use the Sandbox.
Configuration
| Variable | Meaning |
|---|---|
LUX_AI_API |
API key (when no account is signed in) |
LUX_CODE_MOUSE |
0 = classic inline view without mouse handling |
LUX_CODE_MAX_MESSAGES, LUX_CODE_MAX_BODY |
LuxAI request limits for summarizing (also agent/luxai-limits.json) |
LUX_CODE_NAME |
name of the settings folder (~/.config/<name>), default lux-code |
See also
- Lux Code and LSL IDE – full reference including LSL IDE
- LuxAI, LuxAI account
- LSL overview