JoinAI-Agent
A General AI Agent That Thinks and Acts
Systematically building a complete agent capability closed loop from perception, planning, execution to collaboration for complex task scenarios.
中文 • Framework Performance • Project Architecture • Quick Start • Examples
JoinAI-Agent is born from China Mobile's self-developed Jiutian·Juzhi Intelligent Agent Platform. As an AI agent that "thinks and acts", JoinAI-Agent pioneers the definition of next-generation enterprise-level agent engine core capabilities, systematically building a complete agent capability closed loop from perception, planning, execution to collaboration for complex task scenarios, marking an important breakthrough for China Mobile in the field of general agent systems. The open source of JoinAI-Agent will break down technical barriers, accelerate industry collaboration, and continuously empower intelligent upgrades across industries.
This time we have open-sourced the standard framework and basic capabilities of JoinAI-Agent. The entire system adopts a frontend-backend separation architecture, with the frontend supporting file uploads, sandbox task tracking, etc., and the backend based on a DAG execution engine, with built-in multiple professional sub-agents (such as information integration, code expert, report generation, etc.) and a rich tool ecosystem (such as search, file processing, etc.).
In the GAIA benchmark test, JoinAI-Agent achieved a comprehensive score of 90.70, ranking first on the leaderboard.
This open source project open-sources the frontend, agent's react mode, multiple sub-agents (reporter, researcher, coder), various standardized interaction protocols, sandbox, etc.
- End-to-End Complete Product: One-click deployment, out-of-the-box, supporting secondary development and customized extensions
- Standardized Protocol Support: Integrates MCP and A2A protocols, supporting pluggable extensions of tools and agents
- Secure Sandbox Execution: Based on E2B sandbox environment, providing secure code execution capabilities, supporting Shell, file operations, browser automation, etc.
- Flexible Architecture Design: Frontend-backend separation, modular design, easy to integrate and maintain
JoinAI-Agent is a powerful agent engine that integrates the following core capabilities:
- 🔧 Tool Integration: Supports various tools including web search, file operations, code execution, and more
- 🤝 A2A Support: Supports A2A protocol for inter-agent collaboration
- 🔌 MCP Support: Integrates MCP protocol, supporting extensible tools
- 🖥️ Sandbox Execution: Secure code execution environment
- 💬 Web Interface: Simple and easy-to-use frontend interface with file upload support
- 🌐 Browser-Use: Supports browser automation operations, including clicking, typing, scrolling, and other operations, enabling automated web interaction and task execution
- Docker and Docker Compose
- Python 3.12+ (for local development)
- Node.js 20+ (for local development)
- Clone the repository
git clone https://github.com/opencmit/JoinAI-Agent.git
cd JoinAI-Agent- Configure environment variables
Backend service: Copy backend/.env.template to backend/.env, and fill in the necessary keys in the .env file. Model and sandbox configurations are required. E2B_API_KEY can be applied at https://e2b.dev/:
Model examples:
# OpenAI Official API
OPENAI_BASE_URL=https://api.openai.com/v1
OPENAI_API_KEY=sk-your-openai-api-key-here
BASE_LLM=gpt-5# Compatible with other model API services
OPENAI_BASE_URL=https://api.siliconflow.cn/v1
OPENAI_API_KEY=your-api-key
BASE_LLM=deepseek-ai/DeepSeek-V3.2-ExpTo use JINA, SERPER, or Bocha search tools, you need to apply for JINA_API_KEY (https://jina.ai/), SERPER_API_KEY (https://serper.dev), and BOCHA_API_KEY (https://bocha.cn/) and add them to the .env file.
Frontend service: Copy frontend/.env.template to frontend/.env, and fill in the necessary keys in the .env file. The E2B_API_KEY must match the E2B_API_KEY in the backend .env file.
A2A agent: An image parsing agent is preconfigured. To use it, copy backend/a2a_agent/pic_agent/.env.template to backend/a2a_agent/pic_agent/.env, fill in the VL model information in the .env file, and upload images via attachment.
OPENAI_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
OPENAI_API_KEY=sk-your-openai-api-key-here
BASE_LLM=qwen-vl-plus- Start services
docker-compose up -d- Access the application
- Frontend interface: http://localhost:9094
- Backend API: http://localhost:18100
cd backend
pip install -r requirements.txt
# Test backend service
python test_demo.pycd frontend
pnpm install
pnpm devMCP tool configuration is located in backend/config/mcp_server.json. Serper and Jina search tools are already configured. After configuring the corresponding keys, they can be used. Users can configure MCP tools themselves.
A2A agent configuration is in backend/config/a2a_server.json. An image parsing agent is preconfigured. After configuring the corresponding keys, it can be used. Users can configure A2A agents themselves.
|
Browser-Use
Fill in the API key in the browser-use section of browser-use.mp4 |
A2A_Image Parsing
An image parsing agent has been integrated via A2A. Configure VL model information in A2A_agent.mp4 |
|
Report Writing
Enter a report topic, fill in Serper and Jina keys in default.mp4 |
Attachment Processing Supports uploading attachments in various formats. The agent can read and process attachment content, including text files, images, etc. Text attachments are converted to markdown format before subsequent tasks. default.mp4 |
- Sandbox environment is enabled by default to ensure secure code execution
- Sensitive configurations are managed through environment variables
- Do not commit API keys to the code repository
- Plan reasoning mode
- More tool support
- Multi-agent memory
- Rich execution environments
This project is licensed under the Apache License 2.0 open source license.
Note: This project is under active development. APIs and configurations may change. Please refer to the latest documentation and code updates.
