TinyFlow is a lightweight multi-agent orchestration engine built on LangGraph. It routes user queries through four distinct execution modes -- from instant answers to parallel research -- with real-time streaming, persistent memory, and a skill plugin system.
- ⚡ 4 Execution Modes -- Flash (direct answer), Thinking (deep reasoning), Pro (plan & execute), Ultra (parallel research)
- 🤖 ReAct Agent -- Observe-Think-Act loop with multi-tool support (web search, skill invocation)
- 🧩 Graph-based Orchestration -- LangGraph-powered state machine with configurable middleware pipeline
- 📡 Real-time Streaming -- Server-Sent Events with 12+ event types and per-thread independent connections
- 🧠 Memory System -- Fact extraction, confidence scoring, time-decay merging, and context injection
- 🔌 Skill System -- YAML frontmatter registry with keyword matching and hot-reload
- 🎨 Artifact Panel -- Live HTML/CSS/JS preview for generated code and visualizations
- 🧪 302 Tests -- 272 backend (pytest) + 30 frontend (vitest)
 Flash -- Instant answers with rich markdown |
 Pro -- Plan, execute, and generate artifacts |
 Pro + Skill -- Pulse daily tech news with live preview |
 Ultra -- Parallel deep research with report |
TinyFlow generates production-ready artifacts — from interactive slides to research reports — all rendered live in the built-in preview panel.
 Interactive Slides -- AI Agent 2026 trend report, auto-generated presentation |
 Pulse Daily News -- AI & tech daily digest with stats, categories, and trend highlights |
 Deep Research Report -- React vs Vue vs Svelte comprehensive benchmark analysis |
User
|
v
+------------------+ SSE Stream +------------------+
| Frontend | ◀───────────────── | API Gateway |
| Next.js 16 | ──── REST ────────▶| FastAPI |
| React 19 | | /chat /threads |
+------------------+ +--------+---------+
|
+------v-------+
| Graph Engine |
| (LangGraph) |
+------+-------+
|
+----------+----------+----------+---------+
| | | | |
Router Planner Dispatch Executor Reflector
| | | | |
| | | +-----+-----+ |
| | | | ReAct Agent| |
| | | | (Observe → | |
| | | | Think → | |
| | | | Act loop) | |
| | | +-----+-----+ |
+-----+----+----------+---------+---------+
|
+----------+-----------+
| | |
Memory Middleware Skills
Engine Chain Registry
| | |
v v v
Facts Loop/TODO YAML Plugins
Store Detection (hot-reload)
Prerequisites
- Python 3.12+
- Node.js 20+ with pnpm
- uv (Python package manager)
- At least one LLM API key (MiniMax, GLM, OpenAI, or Anthropic)
1. Clone the repository
git clone https://github.com/venaissance/tiny-flow.git
cd tiny-flow2. Backend setup
cd backend
# Copy environment file and add your API key
cp .env.example .env
# Edit .env — at minimum, set one of: MINIMAX_API_KEY, GLM_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY
# Install dependencies and start the dev server
uv sync
make dev
# API server starts at http://localhost:80003. Frontend setup (new terminal)
cd frontend
# Install dependencies
pnpm install
# Start the dev server
pnpm dev
# Open http://localhost:3000tiny-flow/
├── backend/
│ ├── app/
│ │ └── gateway/ # FastAPI application and route handlers
│ │ ├── app.py # ASGI entry point
│ │ └── routers/ # /chat, /threads endpoints
│ ├── core/
│ │ ├── graph/ # LangGraph state machine
│ │ │ ├── builder.py # Graph construction with middleware
│ │ │ ├── state.py # GraphState TypedDict
│ │ │ └── nodes/ # router, plan, dispatch, execute, merge, reflect, respond, skill
│ │ ├── executor/ # ReAct Agent with multi-tool support + ExecutorPool
│ │ ├── memory/ # Fact extraction → scoring → merging → injection
│ │ ├── middleware/ # Loop detection, TODO tracking, context compaction
│ │ ├── models/ # Multi-provider LLM factory (OpenAI, Claude, GLM, MiniMax)
│ │ ├── skills/ # YAML skill registry and router
│ │ └── tools/ # Web search, skill invocation
│ ├── skills/ # Built-in skill definitions (YAML frontmatter)
│ ├── tests/ # 272 pytest tests
│ ├── config.yaml # Model, executor, memory, skill configuration
│ ├── Makefile # install / dev / test / clean
│ └── pyproject.toml
├── frontend/
│ ├── src/
│ │ ├── app/ # Next.js pages and API routes
│ │ ├── components/
│ │ │ ├── chat/ # Chat UI, message rendering, mode selector
│ │ │ ├── inspector/ # Artifact preview panel
│ │ │ └── ui/ # Shared UI components (Radix-based)
│ │ ├── hooks/ # SSE streaming, thread management
│ │ ├── core/ # Config, streaming markdown parser
│ │ └── lib/ # Utilities
│ ├── e2e/ # Playwright E2E tests
│ ├── package.json
│ └── vitest.config.ts
└── docs/
└── assets/ # Screenshots
All backend configuration is in backend/config.yaml:
model:
default: "MiniMax-M2.7" # Default LLM model
providers: # Available LLM providers
- name: openai
- name: claude
- name: glm
- name: minimax
executor:
scheduler_workers: 3 # Parallel task scheduling
execution_workers: 3 # Concurrent task execution
default_timeout: 300 # Task timeout (seconds)
memory:
token_budget: 500 # Max tokens for memory injection
decay_days: 30 # Confidence decay period
min_confidence: 0.7 # Minimum fact confidence threshold
max_facts: 50 # Maximum stored facts
skills:
dirs: [skills/] # Skill definition directories
hot_reload: true # Watch for skill file changes
graph:
max_iterations: 3 # Maximum plan-execute loop iterations# Backend unit & integration tests (272 tests)
cd backend
make test
# Frontend unit tests (30 tests)
cd frontend
pnpm test:run| Category | Technology | Purpose |
|---|---|---|
| LLM Orchestration | LangGraph, LangChain | State machine, ReAct agent, model abstraction |
| Backend Framework | FastAPI, Uvicorn | REST API, SSE streaming |
| Frontend Framework | Next.js 16, React 19 | App router, server components |
| Styling | Tailwind CSS 4 | Utility-first CSS |
| UI Components | Radix UI, Motion | Accessible primitives, animation |
| Code Editor | CodeMirror 6 | Artifact code editing |
| State Management | TanStack Query | Server state synchronization |
| Testing | pytest, vitest, Playwright | Backend, frontend, E2E |
| Package Management | uv, pnpm | Fast, deterministic installs |
| Language | Python 3.12, TypeScript 5.8 | Type safety on both ends |
- Fork the repository
- Create a feature branch:
git checkout -b feat/your-feature - Write tests for your changes
- Ensure all tests pass:
make test(backend) andpnpm test:run(frontend) - Submit a pull request
TinyFlow 是一个基于 LangGraph 的轻量级多智能体编排引擎。它将用户请求智能路由到四种执行模式——从即时回答到并行研究——同时提供实时流式推送、持久化记忆和技能插件系统。
- ⚡ 四种执行模式 -- Flash（直接回答）、Thinking（深度推理）、Pro（规划执行）、Ultra（并行研究）
- 🤖 ReAct Agent -- Observe-Think-Act 循环，支持多工具调用（网页搜索、技能调用）
- 🧩 图编排引擎 -- 基于 LangGraph 的状态机，支持可配置的中间件管道
- 📡 实时流推送 -- SSE 支持 12+ 事件类型，每个会话独立连接，支持并发执行
- 🧠 记忆系统 -- 事实提取 → 置信度评分 → 时间衰减合并 → 上下文注入
- 🔌 技能系统 -- YAML frontmatter 注册，关键词匹配，文件变更热重载
- 🎨 产物面板 -- 生成的 HTML/CSS/JS 代码实时预览
- 🧪 302 个测试 -- 272 后端（pytest）+ 30 前端（vitest）
|
Flash -- 即时回答，丰富 Markdown 排版 |
Pro -- 任务规划、分步执行、产物生成 |
|
Pro + 技能 -- Pulse 科技日报，实时预览 |
Ultra -- 并行深度研究与报告生成 |
TinyFlow 可生成高质量产物——从交互式幻灯片到深度研究报告——均在内置预览面板中实时渲染。
|
交互式幻灯片 -- AI Agent 2026 趋势报告，自动生成演示文稿 |
|
Pulse 科技日报 -- AI 与科技日报，含数据统计、分类聚合、趋势高亮 |
|
深度研究报告 -- React vs Vue vs Svelte 综合性能基准分析 |
用户
|
v
+------------------+ SSE 事件流 +------------------+
| 前端 | ◀───────────────── | API 网关 |
| Next.js 16 | ──── REST ────────▶| FastAPI |
| React 19 | | /chat /threads |
+------------------+ +--------+---------+
|
+------v-------+
| 图引擎 |
| (LangGraph) |
+------+-------+
|
+----------+----------+----------+---------+
| | | | |
路由器 规划器 调度器 执行器 反思器
| | | | |
| | | +-----+-----+ |
| | | | ReAct Agent| |
| | | | (观察 → | |
| | | | 思考 → | |
| | | | 行动 循环) | |
| | | +-----+-----+ |
+-----+----+----------+---------+---------+
|
+----------+-----------+
| | |
记忆引擎 中间件链 技能注册
| | |
v v v
事实存储 循环/TODO YAML 插件
检测 (热重载)
环境要求
- Python 3.12+
- Node.js 20+，安装 pnpm
- uv（Python 包管理器）
- 至少一个 LLM API Key（MiniMax、GLM、OpenAI 或 Anthropic）
1. 克隆仓库
git clone https://github.com/venaissance/tiny-flow.git
cd tiny-flow2. 启动后端
cd backend
# 复制环境变量文件，填入 API Key
cp .env.example .env
# 编辑 .env，至少设置一个：MINIMAX_API_KEY、GLM_API_KEY、OPENAI_API_KEY、ANTHROPIC_API_KEY
# 安装依赖并启动开发服务器
uv sync
make dev
# API 服务运行在 http://localhost:80003. 启动前端（新开终端）
cd frontend
# 安装依赖
pnpm install
# 启动开发服务器
pnpm dev
# 打开 http://localhost:3000tiny-flow/
├── backend/
│ ├── app/
│ │ └── gateway/ # FastAPI 应用和路由处理
│ │ ├── app.py # ASGI 入口
│ │ └── routers/ # /chat, /threads 端点
│ ├── core/
│ │ ├── graph/ # LangGraph 状态机
│ │ │ ├── builder.py # 图构建（含中间件绑定）
│ │ │ ├── state.py # GraphState 类型定义
│ │ │ └── nodes/ # router, plan, dispatch, execute, merge, reflect, respond, skill
│ │ ├── executor/ # ReAct Agent（多工具支持）+ ExecutorPool
│ │ ├── memory/ # 事实提取 → 评分 → 合并 → 注入
│ │ ├── middleware/ # 循环检测、TODO 跟踪、上下文压缩
│ │ ├── models/ # 多 Provider LLM 工厂（OpenAI、Claude、GLM、MiniMax）
│ │ ├── skills/ # YAML 技能注册中心和路由器
│ │ └── tools/ # 网页搜索、技能调用
│ ├── skills/ # 内置技能定义（YAML frontmatter）
│ ├── tests/ # 272 个 pytest 测试
│ ├── config.yaml # 模型、执行器、记忆、技能配置
│ ├── Makefile # install / dev / test / clean
│ └── pyproject.toml
├── frontend/
│ ├── src/
│ │ ├── app/ # Next.js 页面和 API 路由
│ │ ├── components/
│ │ │ ├── chat/ # 聊天 UI、消息渲染、模式选择
│ │ │ ├── inspector/ # 产物预览面板
│ │ │ └── ui/ # 共享 UI 组件（基于 Radix）
│ │ ├── hooks/ # SSE 流式推送、会话管理
│ │ ├── core/ # 配置、流式 Markdown 解析器
│ │ └── lib/ # 工具函数
│ ├── e2e/ # Playwright 端到端测试
│ ├── package.json
│ └── vitest.config.ts
└── docs/
└── assets/ # 截图
后端配置集中在 backend/config.yaml：
model:
default: "MiniMax-M2.7" # 默认 LLM 模型
providers: # 可用 LLM 提供商
- name: openai
- name: claude
- name: glm
- name: minimax
executor:
scheduler_workers: 3 # 并行任务调度数
execution_workers: 3 # 并发任务执行数
default_timeout: 300 # 任务超时（秒）
memory:
token_budget: 500 # 记忆注入的 token 预算
decay_days: 30 # 置信度衰减周期
min_confidence: 0.7 # 最低事实置信度阈值
max_facts: 50 # 最大存储事实数
skills:
dirs: [skills/] # 技能定义目录
hot_reload: true # 监听技能文件变更
graph:
max_iterations: 3 # 规划-执行循环最大迭代次数# 后端单元与集成测试（272 个）
cd backend
make test
# 前端单元测试（30 个）
cd frontend
pnpm test:run| 类别 | 技术 | 用途 |
|---|---|---|
| LLM 编排 | LangGraph, LangChain | 状态机、ReAct Agent、模型抽象层 |
| 后端框架 | FastAPI, Uvicorn | REST API、SSE 流式推送 |
| 前端框架 | Next.js 16, React 19 | App Router、Server Components |
| 样式 | Tailwind CSS 4 | 原子化 CSS |
| UI 组件 | Radix UI, Motion | 无障碍基础组件、动画 |
| 代码编辑器 | CodeMirror 6 | 产物代码编辑 |
| 状态管理 | TanStack Query | 服务端状态同步 |
| 测试 | pytest, vitest, Playwright | 后端、前端、端到端 |
| 包管理 | uv, pnpm | 快速确定性安装 |
| 语言 | Python 3.12, TypeScript 5.8 | 前后端类型安全 |
- Fork 本仓库
- 创建特性分支：
git checkout -b feat/your-feature - 为你的改动编写测试
- 确保所有测试通过：
make test（后端）和pnpm test:run（前端） - 提交 Pull Request