多 LLM Agent 协作讨论模拟器 / Multi-LLM Agent Collaborative Discussion Simulator
Polysynth 是一个多 LLM Agent 协作讨论模拟器，让多个 AI 角色围绕同一话题进行结构化多轮讨论，模拟真实团队协作场景，帮助用户从不同角度深入分析问题，获得更全面的结论。
支持模式：
- 六顶思考帽 — 六种思维角色（白帽·事实、红帽·情感、黑帽·批判、黄帽·乐观、绿帽·创意、蓝帽·主持人）多轮协作
- 辩论赛 — 正方四辩 vs 反方四辩，主持人控场，固定四轮
核心亮点：
- 双模式运行：CLI 终端模式 + Web UI 模式共享同一套核心引擎
- 实时流式交互：WebSocket 传输 LLM 流式 token，逐字显示
- 文件上传：支持 txt/md/pdf/docx/xlsx/pptx，AI 提取摘要注入所有角色 prompt
- Agent 工具：DuckDuckGo 搜索，双阶段调用，关键词 AI 优化
- 多模型路由：每个角色独立配置模型（DeepSeek、Kimi、Anthropic 等）
- 供应商管理：通过 Web UI 或 CLI 动态配置 API Key / Base URL
- 会话持久化：SQLite + JSONL 双写，历史会话可浏览、可回放
- API Key 掩码：GET 返回掩码值，PATCH 忽略掩码值防止误覆盖
# 安装依赖
pip install -r requirements.txt
# 交互式模式（缺失参数时提示输入）
python backend/main.py
# 带参数运行
python backend/main.py run --mode six_hat --topic "AI 会取代程序员吗？" --rounds 3
# 查看历史
python backend/main.py list
# 查看/修改配置
python backend/main.py status
python backend/main.py config host --set-name "主持人"
python backend/main.py config participants six_hat --role-key white --set-model "deepseek/deepseek-chat"# 后端
pip install fastapi uvicorn sqlalchemy aiosqlite websockets python-multipart
pip install PyMuPDF python-docx openpyxl python-pptx # 可选：文件解析
# Windows
.venv\Scripts\python -m uvicorn backend.api.main:app --reload --port 8000
# macOS/Linux
.venv/bin/python -m uvicorn backend.api.main:app --reload --port 8000
# 前端
cd frontend
npm install
npm run dev # http://localhost:5173浏览器打开 http://localhost:5173，选择模式、输入话题、点击开始即可实时观看多 Agent 讨论。
| 层级 | 技术 |
|---|---|
| 后端 | Python 3.10+, FastAPI, SQLAlchemy 2.0 (异步), aiosqlite |
| LLM | LiteLLM（DeepSeek、Kimi、Anthropic、OpenAI 兼容） |
| 前端 | React 18, TypeScript, Vite, Tailwind CSS, shadcn/ui, Zustand, TanStack Query |
| 工具 | DuckDuckGo 搜索、文件解析（txt/md/pdf/docx/xlsx/pptx） |
Browser (React) CLI (Terminal)
| |
| WS / REST | Session.run()
v v
FastAPI ---------------+ TerminalOutputHandler
| WebSocket handler |
| RuntimeConfig.from_db() |
v |
Session (Orchestrator) <-----+
|
+-- ModeRunner (six_hat / debate)
| +-- call_llm() via LiteLLM
| +-- StreamEvent yield
|
+-- SQLite (messages, sessions, config)
+-- JSONL (runtime backup)
| 文档 | 说明 |
|---|---|
| docs/architecture.md | 系统架构、数据流、扩展指南 |
| docs/features.md | 功能概览、使用场景、配置参考 |
| docs/api-contract.md | REST API、WebSocket 协议、组件接口 |
首次运行前，创建 backend/config/secrets.json：
{
"deepseek_api_key": "sk-your-deepseek-key",
"kimi_api_key": "sk-your-kimi-key",
"kimi_base_url": "https://api.kimi.com/coding"
}后端首次启动时会自动从 JSON 配置文件 seed 数据库。之后所有配置可通过 Web UI 或 CLI config 命令修改。
MIT License
Polysynth is a multi-LLM Agent collaborative discussion simulator that enables multiple AI roles to engage in structured, multi-round discussions around a single topic. It simulates real team collaboration scenarios, helping users analyze problems from different perspectives and reach comprehensive conclusions.
Supported Modes:
- Six Thinking Hats — Six roles (Facts, Emotion, Critical, Optimistic, Creative, Host) collaborate in multi-round discussions
- Debate — Pro team (4 speakers) vs Con team (4 speakers), moderator controls, fixed 4 rounds
Highlights:
- Dual-Mode Runtime: CLI terminal mode + Web UI mode share the same core engine
- Real-time Streaming: WebSocket transmits LLM streaming tokens; character-by-character display
- File Upload: Supports txt/md/pdf/docx/xlsx/pptx; AI extracts summaries and injects into all roles' prompts
- Agent Tools: DuckDuckGo search with two-phase calling and query optimization
- Multi-Model Routing: Each role independently configurable (DeepSeek, Kimi, Anthropic, etc.)
- Provider Management: Dynamic API Key / Base URL configuration via Web UI or CLI
- Session Persistence: SQLite + JSONL dual-write; historical sessions browsable and replayable
- API Key Masking: GET returns masked keys; PATCH ignores masked values to prevent accidental overwrites
# Install dependencies
pip install -r requirements.txt
# Interactive mode
python backend/main.py
# With arguments
python backend/main.py run --mode six_hat --topic "Will AI replace programmers?" --rounds 3
# List history
python backend/main.py list
# View/modify configuration
python backend/main.py status
python backend/main.py config host --set-name "Host"
python backend/main.py config participants six_hat --role-key white --set-model "deepseek/deepseek-chat"# Backend
pip install fastapi uvicorn sqlalchemy aiosqlite websockets python-multipart
pip install PyMuPDF python-docx openpyxl python-pptx # Optional
.venv\Scripts\python -m uvicorn backend.api.main:app --reload --port 8000
# Frontend
cd frontend
npm install
npm run dev # http://localhost:5173| Layer | Technology |
|---|---|
| Backend | Python 3.10+, FastAPI, SQLAlchemy 2.0 (async), aiosqlite |
| LLM | LiteLLM (DeepSeek, Kimi, Anthropic, OpenAI-compatible) |
| Frontend | React 18, TypeScript, Vite, Tailwind CSS, shadcn/ui, Zustand, TanStack Query |
| Tools | DuckDuckGo search, file parsing (txt/md/pdf/docx/xlsx/pptx) |
Browser (React) CLI (Terminal)
| |
| WS / REST | Session.run()
v v
FastAPI ---------------+ TerminalOutputHandler
| WebSocket handler |
| RuntimeConfig.from_db() |
v |
Session (Orchestrator) <-----+
|
+-- ModeRunner (six_hat / debate)
| +-- call_llm() via LiteLLM
| +-- StreamEvent yield
|
+-- SQLite (messages, sessions, config)
+-- JSONL (runtime backup)
See docs/architecture.md for detailed documentation.
| Document | Description |
|---|---|
| docs/architecture.md | System architecture, data flow, extension guide |
| docs/features.md | Feature overview, usage scenarios, configuration reference |
| docs/api-contract.md | REST API, WebSocket protocol, component interfaces |
Before first run, create backend/config/secrets.json:
{
"deepseek_api_key": "sk-your-deepseek-key",
"kimi_api_key": "sk-your-kimi-key",
"kimi_base_url": "https://api.kimi.com/coding"
}The first backend startup automatically seeds the database from JSON config files. After that, all configuration can be modified via Web UI or CLI config commands.
MIT License