🧠 ChatLLM — AI Super Assistant with RouteLLM

A full-stack chat application with intelligent RouteLLM model routing for both Chat and Agents modes. Automatically selects the optimal Claude model (Haiku, Sonnet, or Opus) based on query complexity — reducing cost while maintaining quality.

• Auto-routing — scores each query 0–100 across 7 dimensions (length, code depth, reasoning, math/science, planning, creativity, simplicity)
• Three tiers — ⚡ Haiku (fast/cheap), ⚖️ Sonnet (balanced), 💎 Opus (most capable)
• Live preview — shows predicted model as you type, updates in real time
• Route badge on every AI message with complexity bar and score
• Decision callout — explains why a model was selected
• Cost savings tracker — estimates savings vs always using Opus
• Manual override — turn off RouteLLM to force a specific model

• Full conversation history with localStorage persistence
• SSE streaming with real-time token delivery
• Multi-turn context maintained across messages
• Regenerate any response
• Copy individual messages or full chat

• 🔬 Researcher — deep analysis, citations, structured summaries
• 💻 Code Assistant — write, debug, review & explain code
• ✍️ Writer — creative & professional writing
• 📊 Data Analyst — data insights & interpretation
• 📋 Task Planner — goals → actionable steps

Each agent has a specialized system prompt. RouteLLM runs for every agent request with a 3-step reasoning panel (analyze → route → invoke).

Live counters: Haiku / Sonnet / Opus usage, total routed messages, estimated savings vs Opus, average complexity score.

git clone <your-repo>
cd chatllm
npm run install:all

cp server/.env.example server/.env
# Edit server/.env and add your ANTHROPIC_API_KEY

npm run dev

• Frontend: http://localhost:5173
• Backend API: http://localhost:3001

npm run build    # Builds React into server/public/
npm start        # Serves everything from Express

chatllm/
├── package.json              # Root workspace (scripts + concurrently)
│
├── server/
│   ├── package.json
│   ├── .env.example
│   └── src/
│       ├── index.js          # Express server
│       ├── routellm.js       # RouteLLM scoring & routing engine
│       ├── agents.js         # Agent definitions & message builder
│       ├── middleware/
│       │   └── auth.js       # API key validation
│       └── routes/
│           ├── chat.js       # POST /api/chat (SSE streaming)
│           ├── agent.js      # POST /api/agent/:id/chat (SSE + steps)
│           └── route.js      # POST /api/route/analyze
│
└── client/
    ├── package.json
    ├── vite.config.js
    ├── index.html
    └── src/
        ├── main.jsx          # React entry
        ├── App.jsx           # Root component + send logic
        ├── store/
        │   └── index.js      # Zustand store (convs, agents, stats)
        ├── hooks/
        │   └── useRouteLLM.js # SSE streaming hooks
        ├── components/
        │   ├── Sidebar.jsx   # Navigation (Chat + Agents)
        │   ├── Topbar.jsx    # Title bar + actions
        │   ├── StatsBar.jsx  # Live RouteLLM stats
        │   ├── Message.jsx   # Message row with route badge
        │   ├── InputBar.jsx  # Textarea + route controls
        │   ├── EmptyState.jsx # Empty state + suggestions
        │   └── RouteUI.jsx   # RouteBadge, RouteDecision, AgentSteps
        ├── styles/
        │   └── globals.css
        └── utils/
            └── index.js

Stream a chat completion with RouteLLM routing.

Body:

{
  "messages": [{ "role": "user", "content": "..." }],
  "forceModel": null,
  "stream": true
}

SSE Events: routing, token, done, error

Stream an agent response with step-by-step reasoning events.

Agents: researcher, coder, writer, analyst, planner

SSE Events: agent, routing, step, token, done, error

Analyze query complexity without calling Anthropic.

Body: { "text": "..." }

Response:

{
  "score": 72,
  "modelKey": "opus",
  "model": { "id": "...", "label": "Opus 4", "costPer1k": 0.015 },
  "reason": "High complexity query — most capable model selected",
  "isManual": false
}

Returns all model definitions and routing thresholds.