Huiyu Pi,AI coding agent, stripped to essentials.

Huiyu — Tue, 02 Jun 2026 19:04:27 +0000

🇨🇳 中文 · 🇯🇵 日本語 · 🇰🇷 한국어 · 🇪🇸 Español · 🇫🇷 Français · 🇩🇪 Deutsch · 🇧🇷 Português · 🇷🇺 Русский · 🇸🇦 العربية

~80 tokens ~0.3s 90%+ cheaper Local License Stars Forks Issues Release

AI coding agent, stripped to essentials.
~80 tokens system prompt · ~0.3s first token · 4 basic tools · 100% local

Huiyu Pi webUI, A local open-source Agent tool that lets you build your own Harness system from scratch. Built on Pi and pi-forge, compared to IDE tools like Codex and Claude Code, context is reduced to nearly 0 with massive speed improvements. Build cleanly without platform restrictions.

Huiyu Pi Screenshot
Huiyu Pi Demo
Session Management Integrated Terminal
File Browser + Editor Git Integration
Why ~80 Tokens?
Most AI coding tools pack 15,000–28,000 tokens into every request — rules, tool definitions, role prompts, output formatting. The AI spends most of its attention reading boilerplate instead of solving your problem.

Huiyu Pi takes the opposite approach: strip everything non-essential. 4 basic tools. Clean canvas. No baggage.

Before (typical) Huiyu Pi
System prompt overhead 15K–28K tokens ~80 tokens
First token response 2–10 seconds ~0.3 seconds
Per-request cost 0.02–
0.10+ 90%+ cheaper
Tool definitions 10–24+ 4 essentials
Client type Heavy desktop / Electron Pure Web UI
Data privacy Cloud or hybrid 100% local
Why Choose Huiyu Pi?
Built on pi and pi-forge, fixing their lack of a frontend WebUI and clunky interaction details. Huiyu Pi is a browser-based webui for the pi coding agent — it's blazing fast and feels great to use — that's why I'm sharing it.

Advantage Details
⚡ Faster Performance Default context and prompts compressed from ~20K tokens to near zero. AI response time is dramatically shorter.
💰 Lower Token Consumption Most unused context stripped away, drastically reducing per-request API costs.
🎯 Less Context, More Focus Less context = AI stays focused on core instructions for more precise execution.
🔒 Local Deployment = Safe Fully local — API keys and data never leave your machine. Zero data leak risk.
🏗️ Build Your AI Empire Build your own Harness and Agent from scratch. Full control, fully customizable.
🛠️ Fixing the Original's Gaps Fixes pi's lack of WebUI and pi-forge's interaction issues. Blazing fast, incredibly smooth.
Quick Start
npx huiyu-pi
Open http://localhost:9144 in your browser, go to Settings → Providers, enter your API key, and start chatting.

Also available as global npm install, manual clone, or platform scripts — see Installation below.

Features
🔐 Self-hosted & Private
Your code, API keys, and conversation history stay on your own machine. No cloud, no third party, no data leakage.

🧠 Multi-LLM Support
Works with Anthropic Claude, OpenAI GPT/o1/o3, DeepSeek, Google Gemini, Mistral, Groq, xAI, OpenRouter, and more — including local models.

📁 Full File Management
Built-in file browser with CodeMirror editor supporting 10+ languages. Create, edit, search files right in the browser.

🖥️ Integrated Terminal
Full terminal emulator via xterm.js + WebSocket. Multiple tabs, reconnect support, resizable layout.

🔀 Git Integration
View diffs, stage changes at hunk level, explore commit history with git-graph — all from the browser.

🔌 MCP Protocol Support
Connect external MCP servers and expose their tools to your coding agent. Supports both global and project-level configurations.

📱 Mobile-friendly + PWA
Responsive design works on iOS/Android. Install as a PWA for a native-like experience.

🎨 Fully Customizable Theme
Dark and light themes controlled by CSS variables. Create your own skin without rebuilding.

Installation
Prerequisites: Node.js ≥ 20 (download). Build tools required for terminal support: install Visual Studio Build Tools (Windows), xcode-select --install (macOS), or build-essential (Linux). The chat and file browser work without build tools — only the terminal tab requires them.

Option A: npx (no install, runs once)
npx huiyu-pi
Downloads and runs the latest version. Subsequent npx huiyu-pi calls use the cached version. To force an update: npx huiyu-pi@latest.

Option B: Global install (faster subsequent launches)
npm install -g huiyu-pi
huiyu-pi # start server
huiyu-pi --help # show all flags
huiyu-pi --port 4000 # custom port
huiyu-pi --workspace-path ~/Code # custom workspace
To update: npm update -g huiyu-pi. To uninstall: npm uninstall -g huiyu-pi.

Option C: Platform scripts (clone + run)
Clone the repo, then run the start script for your platform:

Platform Command
Windows Double-click start.bat, or run in PowerShell: .\start.bat
macOS / Linux Open Terminal, cd into the repo, then bash start.sh
The start script handles everything: npm install → build server → build client → start. First run takes a few minutes; subsequent runs are fast.

LAN access (share on your local network):

Platform Command
Windows .\start-lan.bat
macOS / Linux bash start-lan.sh
Or via flag: huiyu-pi --host 0.0.0.0

⚠️ Security: Binding to 0.0.0.0 exposes the agent's shell and filesystem to everyone on your network. Only use on trusted private networks.

Option D: Manual (development)
git clone https://github.com/huiyu9144/Huiyu-Pi.git
cd Huiyu-Pi
npm install # install all dependencies (first time may take a few minutes)
npm run dev # start dev server with HMR (http://localhost:9145)
For production builds:

npm run build # build server + client
npm run start # start production server on port 9144
Option E: Docker (recommended for production)
Pull from registry (no clone needed):

docker run -d \
--name huiyu-pi \
-p 9144:9144 \
-v ~/.pi/agent:/home/pi/.pi/agent \
-v ~/.huiyu-pi:/home/pi/.huiyu-pi \
-v $(pwd)/workspace:/workspace \
ghcr.io/huiyu9144/huiyu-pi:latest
Or build locally (for customization):

git clone https://github.com/huiyu9144/Huiyu-Pi.git
cd Huiyu-Pi/docker
cp .env.example .env
docker compose up -d --build
Open http://localhost:9144 in your browser.

Custom UID/GID (fix permission issues on bind mounts):

docker compose build --build-arg PUID=$(id -u) --build-arg PGID=$(id -g)
docker compose up -d
For detailed Docker configuration, see docker/README.md.

API Key Setup
Huiyu Pi manages provider API keys through the Settings UI and stores them in ~/.pi/agent/auth.json (shared with the pi CLI if installed). Keys are never exposed to the browser — the server holds them in memory and proxies all LLM requests.

Via Settings UI (Recommended)
Open http://localhost:9144
Go to Settings → Providers
Select your provider (Anthropic, OpenAI, DeepSeek, etc.)
Paste your API key and save
Supported providers
Anthropic Claude, OpenAI GPT/o1/o3, DeepSeek, Google Gemini, Mistral, Groq, xAI, OpenRouter, and any OpenAI-compatible endpoint (vLLM, LiteLLM, Ollama, etc.).

Custom OpenAI-compatible providers
For self-hosted or third-party endpoints, create ~/.pi/agent/models.json:

{
"custom-gateway": {
"protocol": "openai",
"url": "http://localhost:11434/v1",
"models": ["qwen2.5-coder-32b"]
}
}
Then add the API key in Settings → Providers → custom-gateway.

CLI flag (for scripts / CI)
huiyu-pi --api-key @/path/to/api-key.txt
The @ prefix reads the key from a file. Useful for CI pipelines and Docker secrets.

Configuration
All settings can be controlled via CLI flags, environment variables, or config files. CLI flags take priority over env vars, which take priority over config files.

CLI flags
Flag Description Default
--port Server port 9144
--host Bind address 127.0.0.1
--workspace-path Root directory for projects ~/huiyu-pi-workspace
--api-key Static API key (@file syntax supported) —
--ui-password Browser login password —
--jwt-secret JWT signing key (auto-generated if unset) —
--log-level Log level: fatal error warn info debug trace info
--no-expose-docs Hide Swagger UI from the browser docs exposed
--help Show all flags —
Environment variables
Variable Description
PORT Server port
HOST Bind address (0.0.0.0 for LAN)
WORKSPACE_PATH Root directory for projects
API_KEY Static API key
UI_PASSWORD Browser login password
JWT_SECRET JWT signing key
LOG_LEVEL Log level
EXPOSE_DOCS Set to false to hide Swagger UI
FORGE_DATA_DIR Override state directory (default ~/.huiyu-pi/)
Config files
File Purpose
~/.pi/agent/auth.json Provider API keys (managed via Settings UI)
~/.pi/agent/settings.json Agent settings (model, thinking level, etc.)
~/.pi/agent/models.json Custom OpenAI-compatible providers
~/.huiyu-pi/mcp.json Global MCP server configuration
Customization
The theme is controlled by CSS custom properties in packages/client/src/index.css:

:root {
--bg-primary: #0a0a0a;
--accent: #60A5FA;
}

html[data-theme="light"] {
--bg-primary: #ffffff;
--accent: #2563EB;
}
Tech Stack
Layer Technologies
Frontend React 19, TypeScript 6, Vite 8, Tailwind CSS v4, Zustand, CodeMirror 6
Backend Fastify 5, WebSocket, SSE, JWT
Terminal xterm.js + node-pty
Infrastructure GitHub Actions CI/CD
Community
💬 Join our Discord
⭐ Star us on GitHub
🌐 Visit our Website
Contributing
We welcome contributions! See CONTRIBUTING.md for guidelines.

Acknowledgments
Built on top of two open-source projects:

pi-forge by Devin Marks and contributors
pi by earendil-works and contributors
Author
Follow the author on X (Twitter): @huiyu91444

License
MIT — see the upstream projects for their licenses:

pi-forge
pi
Copyright (c) 2026 Huiyu Pi contributors Copyright (c) 2026 Devin Marks and pi-forge contributors Copyright (c) 2026 earendil-works and pi contributors

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

Huiyu — Tue, 02 Jun 2026 15:46:10 +0000

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!

DEV Community: Huiyu

Huiyu Pi,AI coding agent, stripped to essentials.

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