AgentForge

An orchestrable multi-layer agent platform built on LangChain / LangGraph, delivering the core loop of MultiAgent orchestration → autonomous execution → service publishing.

中文文档

Overview

AgentForge is an orchestrable multi-layer agent platform for building, operating and observing complex AI agent systems in production.

Instead of a single chatbot, you get:

• Multi‑tenant workspaces – isolate data, agents and services per tenant with shared infrastructure.
• SubAgent‑based orchestration – visually compose coordinator + executor agents with layered SubAgent hierarchies.
• Pluggable skills – register reusable atomic skills once, bind them to agents and services many‑to‑many.
• Memory layer – long‑term agent memory plus shared memory with vector retrieval.
• Multiple I/O surfaces – chat UI, HTTP API, workflows/triggers, scheduled tasks and IM channels.
• Service publishing – turn agent graphs into versioned services with API keys and OpenAPI specs.
• Knowledge base service – document‑centric retrieval and timelines via a dedicated microservice.

Feature Highlights

• Deep observability

  • Invocation timeline with per‑step status, tool calls, tokens and durations.
  • Per‑agent debug panel showing routed messages, tool inputs/outputs and LLM request payloads.
  • Static execution graph + per‑invocation highlighting of the exact path taken.

• Sandboxed execution

  • Dedicated Sandbox daemon for running skills and workflows in an isolated environment.
  • Full capture of stdin/stdout, logs and artifacts for each run.
  • Designed for slow / side‑effect‑heavy tools without blocking the main agent loop.

• Workflow engine & triggers

  • Flow‑based orchestration with support for cron, event and manual triggers.
  • Triggers can start workflows, call services, or poke existing conversations.
  • First‑class integration with observability so every run is traceable.

• IM & channel integrations

  • IM channels (e.g. Feishu/Lark) mapped to conversations and services.
  • Webhook entrypoints so external systems can talk to your agents.

• Extensible skills

  • System‑level skills for platform capabilities (ontology, services, observability, etc.).
  • Business skills loaded from the skills/ directory with a unified execution contract.
  • Workflow / knowledge‑base / observability features are all exposed as first‑class skills.

Screenshots

• Product overview – one place to manage agents, skills and services.
  

• Agent workspace – browse, search and configure individual agents.
  

• Skill catalog – register, organize and reuse atomic skills across teams.
  

• Multi-agent orchestration flows – design and inspect how coordinator and executors collaborate step by step.
  

• Sub-agent topology – understand layered flows and execution hierarchy at a glance.
  

• Sandbox runs – execute skills and workflows safely with full input/output visibility.
  

• Triggers – connect agents to schedules, events and external systems.
  

• End-to-end observability – inspect each invocation, timeline and tool call in a unified debug panel.
  

• Per-agent debug info – drill into a single agent’s decisions, tool calls and messages.
  

• IM channels – plug agents into Slack/IM-style channels for multi-user collaboration.
  

Tech Stack

Layer	Technologies
Backend	FastAPI, LangChain / LangGraph, PostgreSQL / SQLite, Redis, ChromaDB / FAISS
Infra & Docs	Docker Compose, Prometheus, VitePress (docs site)

Project Structure

AgentForge/
├── backend/                  # Backend API service
├── knowledge-base-service-py/# Knowledge-base microservice (memvid)
├── sandbox_daemon/           # Sandbox WebSocket daemon
├── skills/                   # Skill script library
├── docker/                   # Docker orchestration
├── docs/                     # Documentation (internal + public)
├── docs-site/                # VitePress documentation site
└── scripts/                  # Utility scripts

See docs/public/README.md for a high-level documentation index.

Using This Repo With an AI Assistant

This project is designed to be friendly to AI coding assistants. If you feed this README.md (plus the referenced files) to an AI agent, it should be able to:

1. Start the full stack with Docker
   • Follow the steps in docker/.env.example to create and adjust docker/.env.
   • Use ./docker/start.sh or docker-compose/make targets as described in this README and in docker/README.md.
2. Run backend and frontend locally (without Docker)
   • Backend entrypoint: backend/app/main.py via uvicorn app.main:app --reload (see backend/README.md).
   • Frontend entrypoint: frontend/src/main.js via npm install && npm run dev in frontend/.
3. Understand core configuration and secrets
   • Runtime settings come from environment variables, primarily configured via docker/.env and parsed in backend/app/core/config.py (Settings class).
   • Production must override JWT_SECRET and DEFAULT_ADMIN_PASSWORD; see validation logic in Settings._require_secrets_in_production.
4. Work with the knowledge base and sandbox daemon
   • Knowledge-base microservice: knowledge-base-service-py/ (HTTP service used by backend when enabled).
   • Sandbox daemon: sandbox_daemon/daemon.py (connects back to backend via --platform-url and --token).
5. Navigate skills and agent logic
   • Skill scripts live in skills/ and are mounted via configuration in backend/app/ (see SKILLS_DIR in config.py).
   • Multi-agent orchestration and execution logic is in backend/app/agents/ and related modules.
6. Find public docs
   • Public, open-source–friendly docs live under docs/public/, including:
     • docs/public/docker/QUICKSTART.md – Docker quickstart.
     • docs/public/topics/auth/README.md – auth and 2FA overview.
     • docs/public/topics/architecture/gateway-pattern.md – selected architecture notes.

When instructing an AI assistant to modify or extend the project, prefer giving it:

• This README.md
• docker/.env.example
• backend/app/core/config.py
• The specific backend/Frontend files you want to change
• Any relevant docs under docs/public/

Quick Start

Prerequisites

• Docker & Docker Compose
• Python 3.10+ (local development)
• Node.js 18+ (for running the docs site locally)

Docker (Recommended)

git clone <repository-url>
cd AgentForge

# Configure environment
cd docker
cp .env.example .env
# Edit .env as needed (API keys, etc.)

# Start all services
./start.sh
# or: docker-compose up -d --build
# or: make up-build

Access:

Service	URL
Backend API	http://localhost:8000
API Docs (Swagger)	http://localhost:8000/docs
Health Check	http://localhost:8000/health

Default admin credentials: admin / admin123 (change DEFAULT_ADMIN_PASSWORD in production).

See docs/public/topics/auth/README.md for 2FA setup.

Local Development

Backend:

cd backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload

Docs site (optional):

cd docs-site
npm install
npm run dev

Core Features

1. Multi‑tenant workspaces

   • Isolated agents, applications and data per tenant.
   • Shared infrastructure layer (DB, vector stores, observability) to reduce operational cost.

2. SubAgent hierarchy

   • Nested SubAgent topology to express multi‑layer task decomposition.
   • Visual configuration of coordinator + executor agents and their responsibilities.

3. Skill plugin & system skill framework

   • Skills are first‑class plug‑ins with a standard skill_execute(**kwargs) contract.
   • Both business skills and platform/system skills share the same registration + execution path.

4. Execution, Sandbox & observability

   • Task decomposition view + live execution timeline, including tool calls and LLM calls.
   • Sandbox daemon for isolated tool execution with full logs and artifacts.
   • Static execution graph with per‑invocation highlighting and Prometheus metrics.

5. Service publishing & workflows

   • Publish agent graphs as standalone API services with API keys and version management.
   • Workflow engine with cron/event/manual triggers, all tied back to conversations and observability.

6. Knowledge base & IM channels

   • Document‑based knowledge base via a dedicated microservice (memvid).
   • IM channels (e.g. Feishu/Lark) mapped into conversations for multi‑user collaboration.

Contributing

See CONTRIBUTING.md for development workflow and guidelines.

License

Apache License 2.0

Security

Please report vulnerabilities via GitHub Security Advisories. See SECURITY.md.

---

AgentForge - 多智能体操作系统

基于 LangChain / LangGraph 构建的多智能体操作系统，聚焦「多租户、多智能体编排、Skill 插件体系、Workflow / Trigger、Sandbox 执行、可观测性、IM 通道与知识库」的完整闭环。

项目概述

相较于单一 Chatbot，AgentForge 更像一套「多智能体操作系统」，提供：

• 多租户工作区：不同租户/环境共享基础设施，又相互隔离应用、服务与数据。
• SubAgent 分层编排：以 SubAgent 拓扑描述协调 Agent 与执行 Agent 的分工与多层任务拆解路径。
• Skill 插件体系：业务 Skill 与系统 Skill 统一注册/发现/执行，复用原子能力，减少重复接入。
• 记忆层：同时支持 Agent 个人长期记忆与共享记忆，基于向量检索。
• 多种入口：Chat UI、HTTP API、工作流触发器（定时、事件、手动）、IM Channel 等。
• 服务化发布：将编排好的智能体能力一键发布为带版本与 API Key 管理的服务，支持 OpenAPI 导出。

能力亮点

• 可观测性优先

  • 会话维度记录每一次调用（同步/流式），构建执行时间线。
  • 每个 Agent / Skill / 工具调用都有独立调试面板：请求体、响应体、Token & 耗时、调用链等一目了然。
  • 静态执行拓扑 +「按次高亮」，支持从调用记录一键高亮本次经过的节点与边。

• Sandbox 执行环境

  • 独立 Sandbox Daemon，通过 WebSocket 与平台通信。
  • 适合 I/O 密集或有强副作用的 Skill，执行过程与日志、产出物完整可追踪。

• 工作流与触发器

  • 基于 Flow 的任务编排，支持 Cron / 事件 / 手动三类 Trigger。
  • 每次触发都对应完整的执行记录和可观测数据，方便排障与审计。

• 知识库与 IM 通道

  • 通过独立微服务提供文档知识库（上传、切分、向量检索、时间线等）。
  • 支持飞书/IM Channel 集成，将 Agent 接入到现有沟通工具中。

技术栈

层次	技术
后端	FastAPI、LangChain / LangGraph、PostgreSQL / SQLite、Redis、ChromaDB / FAISS
基础设施与文档	Docker Compose、Prometheus、VitePress（文档站）

快速开始

前置要求

• Docker & Docker Compose
• Python 3.10+（本地开发）
• Node.js 18+（用于本地运行文档站，可选）

使用 Docker 启动（推荐）

git clone <repository-url>
cd AgentForge

# 配置环境变量
cd docker
cp .env.example .env
# 编辑 .env 文件，配置 API Key 等参数

# 启动所有服务
./start.sh
# 或: docker-compose up -d --build
# 或: make up-build

Compose 项目名已固定为 agentforge，用于避免与其他项目发生服务覆盖。

访问服务：

服务	地址
后端 API	http://localhost:8000
API 文档（Swagger）	http://localhost:8000/docs
健康检查	http://localhost:8000/health

默认管理员： 用户名 admin，密码 admin123（生产环境请修改 DEFAULT_ADMIN_PASSWORD）。

首次登录后可启用两步验证（2FA），详见 docs/topics/auth/README.md。

本地开发

后端：

cd backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload

文档站（可选）：

cd docs-site
npm install
npm run dev

核心功能

1. 多租户工作区

   • 支持按租户/环境隔离应用、服务与数据。
   • 共享底层数据库、向量库与可观测基础设施，降低运维成本。

2. SubAgent 分层编排

   • 通过 SubAgent 拓扑表达协调 Agent 与执行 Agent 的职责边界与多层任务拆解路径。

3. Skill 插件与系统 Skill 框架

   • Skill 以统一的 skill_execute(**kwargs) 协议实现，支持动态加载与注册。
   • 业务 Skill 与系统级 Skill（服务、可观测、知识库等）走同一套执行路径。

4. 执行引擎、Sandbox 与可观测性

   • 任务拆解视图 + 实时执行时间线（含工具调用与 LLM 调用）。
   • 独立 Sandbox Daemon 支持有副作用/长耗时的 Skill 隔离执行，完整保留日志与产出物。
   • 静态执行拓扑 + 按次高亮，并输出 Prometheus 指标。

5. 服务发布与工作流

   • 将编排好的 Agent Graph 发布为带版本与 API Key 管理的独立 API 服务。
   • 工作流引擎支持 Cron / 事件 / 手动触发，并与会话与可观测体系打通。

6. 知识库与 IM 通道

   • 通过独立微服务提供文档知识库能力（上传、切分、向量检索、时间线等）。
   • 支持飞书 / IM Channel 等接入方式，将 Agent 带入实际沟通场景。

开发指南

请参考公开文档入口（docs/public/ 目录）和 CONTRIBUTING.md。

数据存储

• 关系型数据库：Docker 下为 PostgreSQL；本地默认 SQLite（backend/data/agentforge.db）
• 向量数据库：backend/data/chroma/
• Skill 脚本：skills/ 目录
• 日志文件：backend/logs/

贡献指南

见 CONTRIBUTING.md。

许可证

Apache License 2.0

联系方式

通过 GitHub Issues 提问；安全问题请见 SECURITY.md。