🥝 Kivi — Unified AI Chat Interface

Provider-agnostic AI chat with token-level streaming, server-side tools, session persistence, and auto-compaction. One beautiful UI, any backend.

✨ Features

• 🔀 Provider Switching — Switch between OpenAI, vLLM, Copilot, Claude mid-conversation
• ⚡ Token-Level Streaming — Real-time token streaming for all providers (not word-level)
• 🛠️ Server-Side Tools — bash, read, write, edit, glob, grep, web_search, web_fetch
• 💾 Session Persistence — SQLite-backed sessions with full message history
• 📦 Auto-Compaction — Automatically compacts context at 75% of provider's window
• 🎨 Claude-like UI — Dark/light themes, thinking blocks, tool blocks, markdown, code highlighting, LaTeX, SVG preview
• 📊 Token Dashboard — Usage tracking with Plotly charts
• 🔧 Git Dashboard — Built-in diff viewer, commit & push

🚀 Quick Start

pip install kivi-ai
kivi

Open http://localhost:8899 in your browser. That's it.

📦 Install with SDK support

# With GitHub Copilot SDK support
pip install kivi-ai[copilot]

# With Claude Agent SDK support
pip install kivi-ai[claude]

# Everything
pip install kivi-ai[all]

⚙️ Configuration

All configuration via environment variables or CLI flags:

# Set vLLM backend URL
kivi --vllm-url http://your-server:8000

# Custom port
kivi --port 9000

# Or use environment variables
export VLLM_URL=http://your-server:8000
export OPENAI_API_KEY=sk-...
export KIVI_PORT=9000
kivi

CLI Options

kivi                    Start server (default: 0.0.0.0:8899)
kivi --port 9000        Custom port
kivi --host 127.0.0.1   Bind to localhost only
kivi --vllm-url URL     vLLM backend URL
kivi --reload           Dev mode with auto-reload
kivi --help             Show help

🏗️ Architecture

kivi_ai/
├── core/               # Types, interfaces, registry
│   ├── types.py        # Message, StreamChunk, ToolCall, ModelInfo, etc.
│   ├── interfaces.py   # BaseProvider ABC, ToolInterface, SessionStore
│   └── registry.py     # Provider & tool registry singleton
├── providers/          # Provider implementations
│   ├── openai_provider.py   # OpenAI & vLLM (OpenAI-compatible)
│   ├── copilot_provider.py  # GitHub Copilot SDK
│   ├── claude_provider.py   # Claude Agent SDK
│   └── config.py            # Context windows, costs, defaults
├── streaming/          # Stream processing
│   ├── adapter.py      # Normalize streams (filter empty, ensure DONE)
│   └── sse.py          # StreamChunk → SSE text
├── sessions/           # Session management
│   ├── store.py        # SQLite store (async, WAL mode)
│   ├── manager.py      # Session lifecycle & provider switching
│   └── compaction.py   # Auto-compact at 75% context window
├── tools/              # Server-side tool system
│   └── builtins.py     # bash, read, write, edit, glob, grep, web_search, web_fetch
├── frontend/
│   └── index.html      # Claude-like chat UI (single file)
├── server.py           # FastAPI app — unified streaming endpoint
└── cli.py              # CLI entry point

Key Design Decisions

• All providers normalize to AsyncIterator[StreamChunk] — single streaming format
• Sessions are provider-agnostic — switch providers without losing history
• SQLite with WAL mode + ThreadPoolExecutor — non-blocking async I/O
• Atomic message sequencing — no race conditions on concurrent writes
• Single unified endpoint — POST /api/chat/stream handles all providers

🔌 Providers

🌐 API

All endpoints under the server:

📄 License

MIT — see LICENSE