luma

Lightweight coding agent built with Rust. Multi-provider AI support (Anthropic Claude, OpenAI Codex).

Install

macOS, Linux, and WSL (aarch64, x86_64):

curl -fsSL https://raw.githubusercontent.com/nghyane/luma/master/install.sh | sh

Update: luma update

Features

Three modes — switch with Tab:

Mode	Model	Use case
Rush	Haiku	Quick fixes, simple questions
Smart	Opus	Code review, complex problems
Deep	Codex	Advanced analysis, research

@file mention — type @ in prompt to autocomplete file paths. File content injected as context when sent. Multiple files supported.

Image paste — paste (Cmd+V / Ctrl+V) when clipboard has an image. Sent as multimodal content to the model.

Tools — Read, Write, Edit, Bash, Grep, Glob, apply_patch. Web search via server-side (Claude) or client fallback.

Skills — compatible with Claude Code skill format:

• .agents/skills/, .claude/skills/ (project-level)
• ~/.agents/skills/, ~/.claude/skills/, ~/.config/luma/skills/ (user-level)

Sessions — /resume to continue last session, /sessions to browse all, /new to start fresh.

Keyboard Shortcuts

Key	Action
Tab	Cycle mode (Rush → Smart → Deep)
Enter	Send message
Alt+Enter	Newline in prompt
Paste	Text paste, or image if clipboard has no text
Ctrl+T	Cycle thinking level
Escape	Abort streaming (press twice)
Ctrl+C	Abort streaming / clear input / quit
↑ ↓	History / dropdown navigation

Auth

Zero-config — reuses credentials from Claude Code and Codex CLI. OAuth tokens auto-refresh.

Configuration

Preferences stored in ~/.config/luma/preferences.json — mode, model, thinking level. Changed through TUI, no manual editing needed.

Debug: LUMA_DEBUG=1 luma → logs to /tmp/luma.log

License

MIT — see LICENSE.