◆ Lun'Ira

intent → code

A multi-agent AI pipeline CLI. Feed the forge a plain-text intent — refined, tested code comes out.

Quick start

npm install -g @crolix-altf4/lunira
lunira               # the forge opens — setup runs automatically on first launch

[!TIP] No API key yet? Type /demo in the forge bar — a full pipeline runs on mock data, no credentials needed.

[!TIP] Easiest entry point: one OpenRouter key gives you access to 200+ models including free tiers. Lun'Ira will suggest it automatically if no provider is configured.

The forge

One plain-text intent enters. Four agents process it in sequence — each focused on a single role, passing only what the next step needs. On failure, the work goes back into the fire.

  "build a REST API to manage users"
           │
           ▼
  ┌──────────────────┐
  │   Product Owner  │  clarifies goal, requirements, constraints
  └────────┬─────────┘
           │  structured requirements
           ▼
  ┌──────────────────┐
  │     Planner      │  architecture, tech stack, task breakdown
  └────────┬─────────┘
           │  architecture + tasks
           ▼
  ┌──────────────────┐
  │    Developer     │  generates complete, runnable files
  └────────┬─────────┘
           │  code files
           ▼
  ┌──────────────────┐
  │   QA Engineer    │  verdict + score + issues
  └────────┬─────────┘
           │  fail? → back to Developer (up to N times)
           ▼
      ✦ artifact sealed

Why four agents instead of one? Context pollution. A single LLM handling PO + architecture + code + QA in one context degrades fast. Lun'Ira splits each responsibility, keeps outputs typed, and gives you full traceability per run.

[!NOTE] "Ira" is the sin of Wrath (Trinity Seven) — the forge that strikes with precision and force. Part of the Lun' ecosystem.

Features

Pipeline

• Automatic model selection per role — each agent gets the model best suited for its task
• QA iteration loop — on fail, issues are fed back to Dev for a retry (--max-iterations)
• Budget guard — abort if total cost exceeds your limit (--budget-usd, --daily-budget-usd)
• Provider fallback — automatic failover to the next configured provider

7 Providers

| Provider | Free tier | Notes | | ---------- | --------- | -------------------------------------- | | Groq | ✓ | Ultra-fast LPU — default for PO and QA | | Gemini | ✓ | 1M context, strong reasoning | | OpenRouter | ✓ | Gateway to 200+ models, one key | | Ollama | ✓ | Local models, no API key | | Claude | — | Best code quality | | OpenAI | — | GPT-4o, o1 | | NVIDIA NIM | ✓ | Hosted NVIDIA models |

Skills & Plugins

Skills are expertise injected into agent prompts — TypeScript conventions, React patterns, Laravel, security rules, REST design, and more. Think of them as the blacksmith's training.

Plugins are tools the agents can wield — write files, read context, web search, run commands, execute code, open GitHub issues. Bridges to the outside world.

Every plugin carries a safety tier — safe, restricted, or dangerous. The first time you activate a restricted or dangerous plugin from the Arsenal screen, Lun'Ira prompts for consent. Grants are saved in ~/.config/lunira/permissions.json so you're only asked once.

Both are declared per-project via lunira.config.json and selectable per-run from the TUI (/arsenal).

[!NOTE] Skills and plugins are community-extensible via npm packages (lunira-skill-*, lunira-plugin-*). Install with lunira install skill <name> or lunira install plugin <name>, and Lun'Ira discovers them automatically.

TUI

• Animated living flame mascot (Calcifer-inspired) on the idle screen
• Slash command navigation — /history, /arsenal, /setup, /demo
• Live pipeline view with forge fire animation per active step
• Tabbed results: Verdict · Artefacts · Stratégie · Diff
• Apply generated files directly to your dungeon ([a] key)
• Run history (the Annales) with re-run support
• RPG lexicon — Incantation, Dungeon, Forge, Blacksmith level, Artifact

Setup

The forge is cold until you light it. On first launch, Lun'Ira opens setup automatically.

lunira setup                              # interactive provider configuration
lunira config set groq.apiKey <key>       # or configure directly
lunira config set openrouter.apiKey <key> # one key, 200+ models

[!TIP] Keys are stored in ~/.lunira/config.json. Environment variables always take precedence — useful for CI or per-project overrides:

GROQ_API_KEY=<key> lunira run "..."

CLI

lunira                                         # interactive TUI
lunira run "build a REST API"                  # fire the forge headlessly
lunira run "..." --dry                         # preview cost — no LLM calls
lunira run "..." --apply                       # write output to current directory
lunira run "..." --skip qa                     # bypass a role
lunira run "..." --model gemini-2.5-pro        # override model
lunira run "..." --provider openrouter         # override provider
lunira run "..." --budget-usd 0.10             # abort above $0.10
lunira run "..." --max-iterations 3            # allow 3 Dev→QA retries
lunira run "..." --json                        # machine-readable output
lunira ask "what does this file do" --file x   # direct question to the LLM
lunira history                                 # browse the Annales
lunira costs                                   # combustible spent
lunira watch ./src --intent intent.txt         # re-forge on file change
lunira catalog                                 # list available skills and plugins
lunira install skill <name>                    # install a lunira-skill-* package
lunira install plugin <name>                   # install a lunira-plugin-* package
lunira init                                    # scaffold a new project

TUI controls

| Key | Action | | ---------- | ----------------------------------------------------- | | / | Slash commands — type to filter autocomplete | | /demo | Run demo pipeline (no API key needed) | | /history | The Annales — browse past runs | | /arsenal | Select skills & plugins for next run | | /setup | Arm the forge — configure API keys | | ↑ ↓ | Navigate | | m | Swap rune (change model on focused step) | | ↵ | Fire the forge | | 1 2 3 4 | Results tabs — Verdict / Artefacts / Stratégie / Diff | | a | Deploy to dungeon (apply files to current directory) | | s | Seal artifacts (save to ./output/<run-id>/) | | r / q | New forge |

Project config

{
  "skills": {
    "all": ["conventional-commits", "typescript-strict"],
    "dev": ["react-css-modules"]
  },
  "plugins": {
    "dev": ["file_write", "read_file"],
    "qa": ["execute_code"]
  }
}

Skills: typescript-strict · react-css-modules · conventional-commits · project-context · laravel-conventions · test-conventions · api-design · i18n · security

Plugins: file_write · read_file · web_search · run_command · list_directory · create_directory · execute_code · github_create_issue

[!WARNING] Lun'Ira generates code. It does not execute it. The QA agent audits for issues but is not a substitute for human review before deploying to production.

Lun ecosystem

| Project | Role | | ---------------------------------------------------------- | --------------------------------------------------------- | | LunIra | AI dev pipeline — intent → code | | LunAcedia | Information infrastructure — events · actions · AI butler | | LunAvaritia | Mobile companion — Android | | LunImago | Imitation learning — gameplay → ONNX policy | | LunAnima | AI companion core — private |

Built by CrOliX-AltF4 · MIT License · © 2026

Where raw intent finds its final form.