I've been frustrated with AI coding tools that load 15K-28K tokens of system prompts before you can even ask a question. The AI spends most of its attention reading the manual, not solving your code.

So I built Huiyu Pi — a self-hosted AI coding agent that starts at ~80 tokens.

What it does:

Browser-based IDE (no heavy Electron app)

~80 tokens system prompt (not 20K)

~0.3s first token response

90%+ cheaper per request

100% local — your code, API keys, conversations never leave your machine

Multi-LLM: Claude, GPT, DeepSeek, Gemini, Mistral, Groq, xAI, OpenRouter

Built-in terminal, file editor, Git integration

PWA support (works on mobile)

How to try:

npx huiyu-pi

Then open http://localhost:9144

Tech stack:

Frontend: React 19, TypeScript, Vite, Tailwind CSS

Backend: Fastify, WebSocket, SSE

Terminal: xterm.js + node-pty

License: MIT

GitHub: https://github.com/huiyu9144/Huiyu-Pi

Would love feedback from the self-hosted community!