Address the last-mile challenge in rapidly developing multi-agent products.

Most existing open-source agent projects are primarily SDKs or frameworks, requiring users to perform additional development and lacking true out-of-the-box usability. In contrast, our open-source JoyAgent-JDGenie is an end-to-end multi-agent product that can directly answer or resolve user queries or tasks. For example, when a user submits a query like“Provide an analysis of recent trends between the US dollar and gold,”JoyAgent-JDGenie can instantly generate a report in web or PPT format.

• Generality and Customization

  • JoyAgent-JDGenie is a versatile multi-agent framework. To customize functionality for new scenarios, users only need to integrate relevant sub-agents or tools into JoyAgent-Genie. To demonstrate its generality, JoyAgent-JDGenie achieved 75.15% accuracy on the GAIA benchmark Validation set and 65.12% on the Test set, outperforming industry-leading products such as OWL (CAMEL), Smolagent (Hugging Face), LRC-Huawei (Huawei), xManus (OpenManus), and AutoAgent (University of Hong Kong).

• Lightweight and Independence

  • Unlike Alibaba’s SpringAI-Alibaba (which relies on the Alibaba Cloud Bailian platform for LLM capabilities) or Coze (dependent on the Volcano Engine platform), our open-source multi-agent product JoyAgent-JDGenie is lightweight and platform-agnostic.

• Comprehensive Open-Source Offering

  • We have fully open-sourced JoyAgent-JDGenie, including its:
    • Frontend and backend
    • Framework and engine
    • Core sub-agents (e.g., Report Generator Agent, Code Agent, PPT Agent, File Agent)
    • For enhanced performance, we welcome users to leverage JoyAgent with fine-tuned models.

Category	agent	Open Sourced?	Fully Open-Sourced Product?	Dependent on Ecosystem?
SDK	SpringAI-Alibaba	Partial	No (SDK only)	Yes (Alibaba Cloud Bailian Platform)
	Coze	Partial	No (Nieo SDK only)	Yes (Volcano Engine Platform)
Framework	Fellow	YES	No (Eko Agent Framework only)	No
	Dify	YES	No (Workflow-focused framework only)	No
	SkyworkAI	YES	No (Agent framework only)	No
	OpenManus	YES	No (Agent framework only)	No
	Owl	YES	No (Agent framework only)	No
	n8n	YES	No (Agent framework only)	no
Protocol	MCP	Yes	No (Protocol only)	no
	A2A	YES	No (Protocol only)	No
	AG-UI	YES	No (Protocol only)	No
Technical Module	memory0	YES	No (Technical module only)	No
	LlamaIndex	YES	No (Technical module only)	No
Product	Our	YES	Yes (End-to-end open-source agent product)	No

Agent	Score	Score_level1	Score_level2	Score_level3	Organization
Alita v2.1	0.8727	0.8868	0.8953	0.7692	Princeton
Skywork	0.8242	0.9245	0.8372	0.5769	天工
AWorld	0.7758	0.8868	0.7791	0.5385	Ant Group
Langfun	0.7697	0.8679	0.7674	0.5769	DeepMind
JoyAgent-JDGenie(Our)	0.7515	0.8679	0.7791	0.4230	Our
OWL	0.6909	0.8491	0.6744	0.4231	CAMEL
Smolagent	0.5515	0.6792	0.5349	0.3462	Huggingface
AutoAgent	0.5515	0.7170	0.5349	0.2692	HKU
Magentic	0.4606	0.5660	0.4651	0.2308	MSR AI Frontiers
LRC-Huawei	0.406	0.5283	0.4302	0.0769	Huawei
xManus	0.4061	0.8113	0.2791	0.0000	OpenManus

This open-source project is based on JoyAgent-JDGenie, publicly releasing the complete product interface, multiple core agent modes (React mode, Plan and Execute mode, etc.), multiple sub-agents (Report Agent, Search Agent, etc.), and multi-agent interaction protocols.

• End-to-End Multi-Agent Product: Ready out-of-the-box with support for secondary development
• Agent Framework Protocols
  • Support for Diverse Agent Design Patterns
  • Multi-Agent Context Management
  • High-Concurrency DAG Execution Engine: Exceptional execution efficiency
• Sub-Agents and Tools
  • Pluggable sub-agents and tools: Pre-configured with various sub-agents and utilities
  • Multiple file export formats: HTML, PPT, Markdown
  • Plan & Tool Call Optimization: Iteratively enhanced via Reinforcement Learning (RL)
  • End-to-End Streaming Responses

• multi-level：work level VS task level
• multi-pattern：plan and executor VS react

Generates novel tools from existing ones instead of creating from scratch (reducing faulty tool generation):

• Implicit Atomization:
  • Automatically decomposes existing tools into atomic sub-tools
  • No need for manual pre-definition of atomic components
• LLM-Driven Reassembly:
  • Dynamically recombines atomic tools via large language models
  • Enables emergent tool creation without human intervention

1.git clone https://github.com/jd-opensource/joyagent-jdgenie.git

2.Manually update the following configurations in genie-backend/src/main/resources/application.yml:
base_url, apikey, model, max_tokens, model_name
Note for DeepSeek users: Set max_tokens: 8192 for deepseek-chat

Manually update the following environment variables in genie-tool/.env_template:
OPENAI_API_KEY, OPENAI_BASE_URL, DEFAULT_MODEL, SERPER_SEARCH_API_KEY
DeepSeek Configuration:Set DEEPSEEK_API_KEY and DEEPSEEK_API_BASE，Configure DEFAULT_MODEL = deepseek/deepseek-chat，
Replace all occurrences of ${DEFAULT_MODEL} with deepseek/deepseek-chat

3.Build the Docker image
docker build -t genie:latest .

4.Launch the Docker container
docker run -d -p 3000:3000 -p 8080:8080 -p 1601:1601 --name genie-app genie:latest

5.Access Genie via browser
Open http://localhost:3000

If you encounter deployment issues, refer to this video tutorial:【5分钟使用deepseek启动开源智能体应用joyagent-genie-哔哩哔哩】 https://b23.tv/8VQDBOK

• jdk17
• python3.11
• python Environment Setup
  • pip install uv
  • cd genie-tool
  • uv sync
  • source .venv/bin/activate

Ultra-detailed guide reference: Step by Step

Directly start all services via shell:

sh check_dep_port.sh # Verify all dependencies and port occupancy
sh Genie_start.sh  # Launch services directly; restart this script after configuration changes (terminate all services with Control+C)

For deployment guidance, refer to the demonstration video:【joyagent-jdgenie部署演示】 https://www.bilibili.com/video/BV1Py8Yz4ELK/?vd_source=a5601a346d433a490c55293e76180c9d

Edit genie-backend/src/main/resources/application.yml to add MCP server URLs (comma-separated):

You can change the front-end request path to the back-end in ui/.env.

mcp_server_url: "http://ip1:port1/sse,http://ip2:port2/sse"

sh start_genie.sh

After integrating the 12306 ticket tool, initiate: "Plan a 7-day trip for 2 people from Beijing to Xinjiang in July and query train tickets" → Genie designs travel itinerary → Invokes MCP tool for ticket queries → Generates final report

Implementing the BaseTool Interface: Declaring Tool Name, Description, Parameters, and Invocation Methods.

/** * Base Tool Interface */publicinterfaceBaseTool {
    StringgetName(); // Tool name
    StringgetDescription(); // Tool description
    Map<String, Object> toParams(); // Tool parameters
    Objectexecute(Objectinput); // Invoke tool
}
// Weather Agent Example
    publicclassWeatherToolimplementsBaseTool {
    @Override
    publicStringgetName() {
        return"agent_weather";
    }

    @Override
    publicStringgetDescription() {
        return"A weather query agent";
    }

    @Override
    publicMap<String, Object> toParams() {
        return"{\"type\":\"object\",\"properties\":{\"location\":{\"description\":\"地点\",\"type\":\"string\"}},\"required\":[\"location\"]}";
    }

    @Override
    publicObjectexecute(Objectinput) {
        return"Today's weather is sunny";
    }
}

Add the following code in com.jd.genie.controller.GenieController#buildToolCollection to integrate custom Agent.

WeatherTool weatherTool = new WeatherTool();
toolCollection.addTool(weatherTool);

sh start_genie.sh

Core Team: Liu Shangkun,Li Yang,Jia Shilin,Tian Shaohua,Wang Zhen,Yao Ting,Wang Hongtao,Zhou Xiaoqing,Liu min,Zhang Shuang,Liuwen,Yangdong,Xu Jialei,Zhou Meilei,Zhao Tingchong,Wu jiaxing, Wang Hanmin,Xu Shiyue,Liu Jiarun

Core Team: JD.com CHO Enterprise Informatization Team (EI)

We welcome all great ideas and suggestions. If you wish to become a project co-builder, you may submit Pull Requests at any time. Whether it's improving products/frameworks, fixing bugs, or adding new features - all contributions are highly valued. Before contributing, you need to read and sign the Contributor Agreement and send it to org.developer3@jd.com. Chinese Version，English Version

For academic references or inquiries, please use the following BibTeX entry:：

@software{JoyAgent-JDGenie,
  author = {Agent Team at JDCHO},
  title = {JoyAgent-JDGenie},
  year = {2025},
  url = {https://github.com/jd-opensource/joyagent-jdgenie},
  version = {0.1.0},
  publisher = {GitHub},
  email = {jiashilin1@jd.com;liyang.1236@jd.com;liushangkun@jd.com;tianshaohua.1@jd.com;wangzhen449@jd.com;yaoting.2@jd.com}
}

Contact Us