PH Agent Hub: Open-source Solution for Multi-tenant AI Platform Based on Microsoft Agent Framework

PH Agent Hub is a modular multi-tenant AI platform built on Microsoft Agent Framework (MAF). It provides a complete chat interface and admin backend, supporting multi-model integration, session management, persistent memory, and one-click Docker deployment, making it suitable for enterprise-level AI application implementation. This project is designed to address core pain points in enterprise AI implementation (such as data isolation across teams, unified model management, persistent session storage, secure access permissions, etc.), aiming to serve as the infrastructure foundation for enterprise AI applications.

Background

With the rapid development of Large Language Model (LLM) technology, enterprises face many challenges in AI integration: data isolation needs during multi-team collaboration, unified management of different AI models, persistent storage of session history, and secure and controllable access permission management. Traditional single-user chat tools can hardly meet enterprise-level scenario requirements.

Project Overview

PH Agent Hub was created and open-sourced by developer Panayiotis Halouvas. Built on Microsoft Agent Framework (MAF), it uses the React + FastAPI tech stack and provides complete AI application infrastructure. It not only has a chat interface for end-users but also includes a fully functional admin backend, supporting one-click Docker containerized deployment to reduce operation and maintenance complexity.

Frontend Architecture

Built with React 18 and TypeScript, using Vite as the build tool, Ant Design 5 for UI component library, and TanStack Query for data state management. It is divided into a chat area (end-user interaction) and an admin backend (operations for admins/tenant managers).

Backend Architecture

Uses Python's FastAPI framework, with SQLAlchemy 2.0 for database operations, and core integration of Microsoft Agent Framework (MAF) to support complex AI workflow orchestration.

Data Layer Design

• MariaDB 11: Stores structured data such as users, sessions, messages, etc.
• Redis 7: Cache and session state management
• MinIO: S3-compatible object storage for file upload and management

End-user features include:

• Real-time Streaming Response: Implements real-time streaming output of AI responses via Server-Sent Events (SSE)
• Multi-model Support: Allows switching between models like DeepSeek, OpenAI, Anthropic, etc. Admins can configure available models for tenants
• Template and Skill System: Predefined prompt templates, supporting personal exclusive prompts; skills encapsulate advanced AI capabilities, combining tools to complete complex tasks
• File Upload and Storage: Files are securely stored in MinIO, supporting scenarios like document analysis
• Memory Management: Persistently remembers user conversation content and preferences; users can view/manage memory entries
• Session Tool Activation: Dynamically activates tools like ERPNext, membrane, etc.
• Message Branching and Feedback: Supports message editing, regeneration, and provides like/dislike feedback
• Temporary Session Mode: Sessions on sensitive topics leave no traces in the database

Admin backend features cover:

• Tenant Management: Create multiple independent tenant environments with complete data isolation
• User Management: Invite users, set roles (admin/manager), reset passwords
• Model Configuration: Add and manage AI model providers, encrypt and store API keys, assign models to tenants
• Tool Configuration: Register ERPNext, membrane, and custom tools
• Usage Analysis: Statistically analyze Token usage by tenant, supporting cost control and capacity planning
• Audit Logs: Record all management operations to meet compliance requirements

DeepSeek Stabilizer

Optimized for DeepSeek models, it automatically handles: inference content stripping, JSON repair, and failure retries to improve stability.

Deployment

• Development Environment: One-click startup via Docker Compose (clone repository → configure env → docker compose up)
• Production Environment: Use Traefik reverse proxy, automatically configure Let's Encrypt SSL certificates, support health checks and external volume mounting

Security Recommendations

• Immediately change the default admin password in the production environment
• Use strong random strings as JWT_SECRET and ENCRYPTION_KEY
• Regularly rotate API keys of AI providers
• Enable audit logs and perform regular backups

PH Agent Hub is suitable for the following scenarios:

• Enterprise Internal AI Platform: Large enterprises deploy it as an internal collaboration platform, with different departments as independent tenants
• AI SaaS Service Provider: Quickly build multi-tenant AI SaaS services based on this
• Development Team Prototype Verification: Use the ready-made user system, session management, and tool expansion capabilities to quickly verify AI application prototypes

PH Agent Hub is a well-designed, fully functional open-source solution for enterprise-level AI platforms. It addresses core implementation pain points such as multi-tenancy, model management, and session persistence. Its architecture is concise and deployment is convenient. The scalability based on Microsoft Agent Framework supports complex AI workflows, and the optimization for domestic models like DeepSeek reflects attention to the local market. For teams looking for an enterprise-level AI platform, whether using it directly or for secondary development, it is a valuable reference and starting point.