Your Local AI Software Engineer
A local-first, multi-agent coding workspace that brings autonomous development capabilities to your desktop.
Three specialized agents — Plan, Code, Deploy — orchestrated through a unified runtime. Each agent is a pluggable prompt model with its own toolset, sharing a common execution loop.
| Multi-Agent Runtime | Decoupled agent engine with CodingAgent, PlanAgent, DeployAgent — each with specialized tools and prompts |
| Live Thinking Stream | SSE-powered real-time visualization of chain-of-thought, tool calls, and terminal outputs |
| Parallel Tool Execution | Read-only tools (search, glob, read) run concurrently via thread pool — faster responses |
| @Skill System | Type @ to inject project-scoped markdown rules, or let the AI auto-activate them |
| Monaco Editor | Full code editor with file tree, diff view, and git integration |
| Real Terminal | xterm.js + PTY backend for real shell interaction with wait/confirm safety |
| Kanban Board | Built-in task tracking with drag-and-drop, rich text descriptions |
| Plan Builder | Interactive Q&A planning with rich text plan editor before coding |
| OpenAI-Compatible | Works with any OpenAI-format API — GPT, DeepSeek, Qwen, local models |
┌──────────────────────────────────────────────────┐
│ Web Dashboard │
│ React 19 · Tailwind v4 · Monaco · xterm.js │
└────────────────────┬─────────────────────────────┘
│ HTTP / SSE
┌────────────────────▼─────────────────────────────┐
│ FastAPI Server │
│ Session Store · Skills Registry │
└──────┬──────────┬──────────┬──────────────────────┘
│ │ │
┌────▼───┐ ┌───▼────┐ ┌──▼─────┐
│ Plan │ │ Coding │ │ Deploy │ Pluggable
│ Agent │ │ Agent │ │ Agent │ Prompt Models
└────┬───┘ └───┬────┘ └──┬─────┘
│ │ │
┌────▼─────────▼─────────▼─────┐
│ Unified Agent Engine │ Shared Runtime
│ Tool Registry · Executor │
└──────────┬───────────────────┘
│
┌──────────▼───────────────────┐
│ OpenAI-Compatible Client │ Any LLM Provider
└──────────────────────────────┘
- Python >= 3.10
- Node.js >= 18 (with pnpm)
Option 1 — One-click (Windows)
.\start.batChoose 2 to install dependencies and launch both servers.
Option 2 — PowerShell script
conda activate base
powershell -ExecutionPolicy Bypass -File .\scripts\quick-start.ps1 -StartBackend -StartFrontendOption 3 — Manual
# Backend
pip install -r requirements.txt
python -m uvicorn fastapi_app.main:app --port 3001
# Frontend
cd frontend && pnpm install && pnpm devcopy .env.example .envSC_AGENT_API_KEY=your-api-key
SC_AGENT_BASE_URL=https://api.openai.com/v1
SC_AGENT_MODEL=gpt-4oWorks with any OpenAI-compatible endpoint — DeepSeek, Qwen, local Ollama, etc. Without .env, the backend runs in Demo Mode with mock replies.
Open http://localhost:8888 and start coding.
Use the agent engine programmatically:
from agent import Agent, AgentLLMConfig, ChatSession, OpenAICompatibleClient
from coding_agent import CodingPromptModel, build_coding_tools
config = AgentLLMConfig.from_env(".env")
client = OpenAICompatibleClient(config)
agent = Agent(
model=CodingPromptModel(client),
tools=build_coding_tools(),
workspace="path/to/project",
)
session = ChatSession(agent=agent)
response = session.ask("Refactor the auth module")
print(response.final_output)Create markdown skill files to inject domain knowledge:
---
name: api-guidelines
description: REST API coding conventions for this project
---
# API Guidelines
- Use snake_case for endpoints
- Always validate with Pydantic v2
...Place at .agents/skills/<name>/SKILL.md (local) or builtin_skills/<name>/SKILL.md (global).
Type @ in the chat to pick a skill, or let the agent auto-discover relevant ones.
Built-in skills include: code-review, frontend-design, security-audit, systematic-debugging, test-driven-development, and more.
| Tool | Description |
|---|---|
list_file |
List directory contents with depth control |
glob_file |
Glob pattern file search |
read_file |
Read file content with pagination |
grep_file |
Regex search across files |
write_file |
Create or overwrite a file |
replace_file |
Safe chunk-based file editing |
execute |
Run shell command with timeout |
terminal_input |
Send input to running terminal |
terminal_wait |
Wait for terminal output |
Read-only tools run in parallel when possible — no waiting for sequential file reads.
SuperCode/
├── agent/ # Core agent framework (runtime loop, events, tool registry)
├── coding_agent/ # Coding agent — file ops, shell, git tools
├── plan_agent/ # Planning agent — task decomposition, checklists
├── deploy_agent/ # Deploy agent — SSH, testing, packaging
├── fastapi_app/ # Backend server (SSE, terminal PTY, session store)
├── frontend/ # React dashboard (Vite, Tailwind v4, Monaco)
├── builtin_skills/ # Pre-packaged markdown skills
├── examples/ # CLI entrypoint and demo workspace
├── scripts/ # Bootstrap and build scripts
└── start.bat # One-click Windows launcher
三种专属智能体 — 规划、编码、部署 — 通过统一运行时编排。每个智能体是可插拔的提示模型，拥有独立工具集，共享执行引擎。
| 多智能体架构 | 解耦引擎，CodingAgent、PlanAgent、DeployAgent 各自专精 |
| 实时思考流 | SSE 驱动，可视化链式推理、工具调用、终端输出 |
| 只读工具并行 | 搜索、读取等只读工具线程池并行执行，响应更快 |
| @技能系统 | 输入 @ 注入项目级 Markdown 规则，或由 AI 自动激活 |
| Monaco 编辑器 | 代码编辑 + 文件树 + Diff 视图 + Git 集成 |
| 真实终端 | xterm.js + PTY 后端，支持交互等待与确认安全机制 |
| 看板管理 | 内置拖拽看板，富文本任务描述 |
| 计划构建器 | 编码前交互式问答规划，富文本计划编辑器 |
| OpenAI 兼容 | 支持任何 OpenAI 格式 API — GPT、DeepSeek、Qwen、本地模型 |
- Python >= 3.10
- Node.js >= 18（需安装 pnpm）
方式一 — 一键启动（Windows）
.\start.bat选择 2 自动安装依赖并启动前后端。
方式二 — PowerShell 脚本
conda activate base
powershell -ExecutionPolicy Bypass -File .\scripts\quick-start.ps1 -StartBackend -StartFrontend方式三 — 手动启动
# 后端
pip install -r requirements.txt
python -m uvicorn fastapi_app.main:app --port 3001
# 前端
cd frontend && pnpm install && pnpm devcopy .env.example .envSC_AGENT_API_KEY=你的密钥
SC_AGENT_BASE_URL=https://api.openai.com/v1
SC_AGENT_MODEL=gpt-4o兼容任何 OpenAI 格式接口 — DeepSeek、Qwen、本地 Ollama 等。未配置 .env 时，后端以 Demo 模式 运行模拟回复。
打开 http://localhost:8888 开始使用。
MIT License — Built with ❤️ by the open-source community
Issues and PRs welcome → GitHub Issues