AgentChat: A FastAPI Multi-Agent Conversation Platform With RAG and MCP

An LLM Platform for Multi-Agent Collaboration, Not a Single-Bot Chat

Single-model chatbot integrations are one approach to AI conversation. AgentChat takes a different position: it is a platform for running multiple agents that can collaborate on complex tasks through multi-round dialogue, with built-in tools, a knowledge base, and workflow support.

The system supports sub-agent cooperation where one agent delegates to others, task automation, and goal-directed workflows. The README describes the primary components as an AI conversation engine, an intelligent agent system, a knowledge base system, and a tool ecosystem with over ten built-in tools plus custom extension support.

The platform is entirely self-hosted. A live demo is available at agentchat.cloud, and the documentation site is hosted on the project's GitHub Pages. The backend is FastAPI running on Python 3.12+, and the frontend is Vue 3.4+ with Element Plus and TypeScript.

Architecture: Separated Frontend, FastAPI Backend, and Storage Dependencies

The repository root contains src/, docker/, docs/, and scripts/ directories. The backend lives in src/backend and the frontend in src/frontend, reflecting the front-back separation the README describes.

The technology stack from the README: backend uses FastAPI, Python 3.12+, LangChain, MySQL, Redis, and ChromaDB. The frontend uses Vue 3.4+, Element Plus, Pinia, Vite 5, and TypeScript. Deployment uses Docker, Docker Compose, Poetry, and npm.

MySQL 8.0+ stores relational data. Redis 7.0+ handles caching and session state. ChromaDB serves as the default vector database for RAG. The system also supports Milvus as an alternative vector database and Elasticsearch for additional search capabilities. Object storage uses either OSS or MinIO, with MinIO being the local option.

All these storage dependencies are external; the Docker Compose configuration manages them as services in docker/docker_config.yaml.

Deploying AgentChat With Docker or Running It Locally

The Docker path is the primary deployment method. Clone the repository, configure the environment file, and start:

# 1. 克隆项目
git clone https://github.com/Shy2593666979/AgentChat.git
cd AgentChat

# 2. 编辑配置文件
vim docker/docker_config.yaml

# 3. 启动
cd docker
docker-compose up --build -d

For a local run without Docker, the README gives a two-part sequence. First, clone the repository and enter it. Then start the backend from src/backend by installing requirements via pip or uv, and start the frontend:

cd src/frontend

# 下载依赖
npm install
npm run dev

After the backend is running, the FastAPI Swagger documentation is available at /docs. Windows users have a one-click deployment script (start_win.bat) added in a later update. Docker 20.10+ is required for the Docker path.

RAG, Three-Layer Memory, and MCP Protocol Integration

AgentChat's knowledge base uses RAG with multiple format support, semantic chunking, and vector retrieval. ChromaDB handles the vector store by default, with Milvus as an alternative.

Memory is a three-layer system introduced in version 2026-4-12. Short-term memory keeps the most recent 3000 tokens of conversation for immediate context continuity. When conversation history exceeds 3000 tokens, the system automatically summarizes it. Long-term memory persists user preferences and habits across sessions for personalization.

MCP (Model Context Protocol) support allows runtime dynamic loading of external MCP servers. The HITL (Human-In-The-Loop) feature enables conversational, step-by-step MCP Server generation from OpenAPI specs. The README describes it as a way to convert API definitions into MCP servers through a dialogue process where users can intervene at key decision nodes rather than accepting an automated output without review.

Tools support a dependency chain calling pattern: if tool C depends on the result of tool B, which depends on tool A, the agent resolves the ordering automatically (A then B then C).

Skills, Sub-Agents, and the Custom Tool Extension Path

Agents can be extended with Skills. A Skill binds progressive prompt loading to an agent, teaching the model how to handle a task type step by step. The README describes this as enabling the model to learn through structured prompts rather than relying solely on its base training.

Custom tools are added by uploading a Swagger or OpenAPI spec through the platform UI, which builds the tool configuration from the API definition. This path avoids writing custom code for tools that already have an OpenAPI spec.

The platform includes a data dashboard that filters call counts and token usage by agent, model, and time range. A workspace feature allows switching between workspace and an application center view.

Default agents are built in, but the platform supports user-defined agents. Multiple agents can run in the same session and collaborate on tasks, with the sub-agent cooperation mechanism distributing work across them.

The LangChain v1.0 Migration and What It Breaks

Starting from AgentChat v2.2.0, the project upgraded LangChain from the 0.x branch to 1.0. The README documents this as a significant breaking change.

Installations at v2.1.x and below use the LangChain 0.x API. From v2.2.0 onward, the LangChain 1.0 API is used, which introduced major changes to tool and agent configuration patterns. The README recommends reviewing the migration guide before upgrading. Projects or forks that customized agent behavior through LangChain 0.x APIs need to update those customizations before moving to v2.2.0+.

The same update also resolved dependency conflicts between Pydantic, LangChain, and FastAPI that affected earlier versions. The pyproject.toml dependency pinning reflects the specific version combinations that are known to work together.

The repository has no GitHub releases, so version history is tracked through the README changelog. The last push was on 2026-08-27.

AgentChat vs Microsoft AutoGen AgentChat

Searching for 'AgentChat' returns significant results for Microsoft AutoGen's AgentChat module, which is a different project. AutoGen AgentChat is Microsoft's Python library for building multi-agent AI applications. It is part of the AutoGen framework and focuses on defining agent types, team structures, and message-passing patterns in Python code.

Shy2593666979/AgentChat is a full-stack application platform with a web UI, a database-backed user and agent management system, and a self-hosted deployment model. It bundles a working Vue frontend and a FastAPI backend rather than being a programming library. The target user is a developer or team that wants to run a multi-agent conversation service without building the UI and backend plumbing themselves.

AutoGen AgentChat's strength is code-level flexibility for developers who want to define agents programmatically and compose them in complex pipelines. AgentChat's strength is a ready-to-deploy platform with UI, user management, RAG, and MCP support included.

Editorial conclusion

AgentChat suits developers who want a self-hosted multi-agent conversation system with built-in RAG, MCP server support, and a ready-made Vue 3 UI, rather than coding those components from scratch. It requires Python 3.12+, MySQL 8.0+, Redis 7.0+, and Node.js 18+. If upgrading from a version below v2.2.0, the LangChain 1.0 migration is a breaking change that requires reviewing the migration guide before updating. Docker Compose is the lowest-friction deployment path.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. Shy2593666979/AgentChat on GitHub