Why LLMTrio?

Each LLM has strengths. Claude reasons deeply. Codex writes fast code. Gemini researches broadly. LLMTrio routes each task to the best model — automatically — and runs them in parallel.

One prompt in, production-ready output out.

Quick Start

npx llmtrio

Or clone and run:

git clone https://github.com/jee599/LLMTrio.git
cd LLMTrio
./trio

The dashboard opens in your browser. Install CLIs, authenticate, and start workflows — all from the UI. No terminal needed after the first command.

How It Works

You → "Build a REST API with auth"
          ↓
    ┌─────────────────────────────────────┐
    │           LLMTrio Engine            │
    │                                     │
    │  Discover → Define → Develop → Deliver
    └──┬──────────┬──────────┬────────────┘
       │          │          │
   Claude 🟠   Codex 🔵   Gemini 🟣
   Architecture  Code       Research
   Code Review   Testing    Documentation

1. Enter a prompt — LLMTrio breaks it into tasks
2. Auto-routes — Each task goes to the best model
3. Parallel execution — All agents work simultaneously
4. Real-time dashboard — Watch progress, costs, and outputs live

Default Model Routing

| Task | Agent | Model | |------|-------|-------| | Architecture | Claude | Opus 4.6 | | Code Review | Claude | Opus 4.6 | | Implementation | Codex | GPT-5.4 | | Testing | Codex | GPT-5.4 | | Research | Gemini | 3.1 Pro | | Documentation | Gemini | 3.1 Pro |

Fully configurable in .trio/routing.json.

Dashboard Features

• Agent Status — Live output, current task, health indicators
• Kanban Board — Drag-and-drop tasks, priorities, dependencies, search & filter
• Cost Tracking — Per-agent and total spend with budget limits
• Model Routing — Change which model handles which task type
• Zero-Terminal Setup — Install & login CLI tools directly from browser

CLI Commands

./trio                        # Open dashboard (default)
./trio start "prompt"         # Start full 4-phase workflow
./trio task claude "prompt"   # Run single agent task
./trio login                  # Check CLI auth status
./trio status                 # Show workflow state
./trio stop                   # Stop all processes

4-Phase Workflow

| Phase | What Happens | |-------|-------------| | Discover | Requirements analysis, tech research | | Define | Architecture design, task breakdown | | Develop | Implementation, testing | | Deliver | Code review, documentation |

Architecture

Browser (dashboard.html)
    ↕ SSE + HTTP
Dashboard Server (localhost:3333)
    ↕ fs.watch + JSON
File System (.trio/)  ←→  Command Watcher  →  CLI Agents

• Zero dependencies — Node.js built-in modules only
• No build step — Single HTML file frontend
• No database — JSON files in .trio/ for all state
• Resilient — Dashboard can close/reopen without losing state

Project Structure

LLMTrio/
├── trio                        # CLI entry point
├── package.json                # npx support
├── scripts/
│   ├── octopus-core.js         # Multi-agent orchestration engine
│   ├── dashboard-server.js     # HTTP + SSE server (Node.js only)
│   ├── dashboard.html          # Frontend (single file, no build)
│   ├── command-watcher.sh      # Command queue processor
│   └── model-checker.sh        # Model update checker
├── config/
│   ├── default-routing.json    # Default task → model routing
│   └── default-budget.json     # Default budget limits
└── .trio/                      # Runtime data (gitignored)

Requirements

• Node.js v18+

That's it. CLI tools (Claude Code, Codex, Gemini CLI) can be installed from the dashboard.

License

MIT