DeerFlow 2.0 - ByteDance's open-source super intelligent agent framework | AI toolkit

DeerFlow 2.0 is ByteDance’s open source super-agent framework, which adopts the “main agent + 11-layer middleware chain + dynamic sub-agent” architecture to achieve multi-agent collaboration through LangGraph. The framework has built-in 10+ pluggable skills such as in-depth research, data analysis, and audio and video creation, and supports Docker/K8s isolation sandbox execution. Natively adapted to Feishu, Telegram, and Slack, it can run without the need for a public IP. Version 2.0 has been upgraded from a single research tool to a general-purpose agent runtime, supporting complex task processing from minutes to hours, and can generate websites, PPTs, comics and other content with one click.

Key features of DeerFlow 2.0

• agent architecture : Adopt a collaboration model of overall planning by the main agent and parallel execution of dynamic sub-agents to achieve efficient decomposition and collaborative processing of complex tasks.
• Pluggable skill system : Built-in more than 10 skill modules such as in-depth research, data analysis, PPT generation, and web design, supporting users to customize and expand new capabilities through MCP interfaces and Python functions.
• Isolate sandbox execution : Provides three operating modes: local, Docker and Kubernetes, and creates an independent container environment for each task based on Byte AIO Sandbox to ensure execution security and resource isolation.
• context engineering : Systematically solve the problem of insufficient context windows in long-term tasks through mechanisms such as automatic summary compression, external storage of intermediate results, and subtask current limiting.
• long term memory : Supports cross-session persistent storage of user portraits, preferences, and accumulated knowledge, and all data is saved locally and fully controlled by the user.
• IM channel integration : Natively adapted to the three major instant messaging platforms Feishu, Telegram and Slack, it can interact with agents through commands without requiring a public IP address.
• Compatible with multiple models : Supports mainstream large models such as OpenAI, Gemini, DeepSeek, and Doubao Seed, and flexibly switches between different underlying capabilities through a unified configuration interface.

Key information and usage requirements for DeerFlow 2.0

• Project background : ByteDance’s open source super-agent framework, using the MIT license.
• Release time : Version 2.0 was officially released on February 28, 2026, and topped the list of GitHub Trending on the day of release.
• core positioning : Comprehensive upgrade from the in-depth research tool of version 1.0 to a general agent runtime (Super Agent Harness).
• mission capability : Supports processing of complex multi-step tasks ranging from minutes to hours, covering a variety of scenarios such as research, coding, and creation.
• Python version : Requires Python 3.12 or higher to be installed on the system.
• Node.js version : Requires the system to have Node.js 22 or higher installed.
• Package manager : Need to install pnpm and uv as front-end and back-end package management tools.

How to use DeerFlow 2.0

• Clone repository :Execute git clone https://github.com/bytedance/deer-flow.git Clone the project locally and enter the project directory cd deer-flow.
• Generate configuration :Run makeconfig Automatically generated config.yaml and .env configuration file.
• Configure model :edit config.yaml file, add the required model (such as GPT-4, Gemini, DeepSeek, Beanbao, etc.), and set parameters such as model name, API type, key variable, etc.
• set key : in .env Fill in the API key of each service provider in the file, such as OPENAI_API_KEY=your-key.
• Choose deployment method Docker way :Execute make docker-init Pull the sandbox image and execute it again make docker-start Start the service.
• local mode :Execute make check Check dependencies and then execute make install Install dependencies and finally execute makedev Start the service. access use : Open the browser and visit http://localhost:2026 to enter the web interface and enter the task instructions to start using it. IM channel access : in config.yaml Open Feishu, Telegram or Slack configuration, set the App ID, key and other parameters of the corresponding platform, and interact with DeerFlow through commands in the chat software.

DeerFlow 2.0 project address

• Project official website ：https://deerflow.tech
• GitHub repository ：https://github.com/bytedance/deer-flow

Comparison of similar competing products of DeerFlow 2.0

Contrast Dimensions	DeerFlow 2.0	JVS Claw (Alibaba)	QClaw (Tencent)
Developer	ByteDance	Alibaba Cloud	Tencent
Open source agreement	MIT (fully open source)	Closed source	Closed source
core architecture	Main agent + 11-layer middleware chain + dynamic sub-agent	Based on OpenClaw packaging	Based on OpenClaw minimalist packaging
Deployment method	Docker/local/K8s, supports privatization	You need to apply for an invitation code and use it in the cloud	Need to apply through Tencent channels
Sandbox mechanism	Byte AIO Sandbox, three-level isolation	Inherit the OpenClaw sandbox	Inherit the OpenClaw sandbox
Skill expansion	Pluggable Skill system supports customization	Preset skills + custom extensions	Preset skills + inspiration square
IM integration	Native support for Feishu, Telegram, and Slack	No native IM support yet	WeChat Mini Program “QClaw Butler”
memory ability	long-term memory, local storage	Inherit OpenClaw memory mechanism	Inherit OpenClaw memory mechanism

Application scenarios of DeerFlow 2.0

• in-depth research : Automatically collect multi-source information and generate research reports, suitable for complex information processing tasks such as academic research, competitive product analysis, and industry research.
• web development : Generate complete and deliverable website pages with one click, automating the entire process from UI design to front-end code, such as football league official website, 3D interactive weather interface, etc.
• content creation : Transform complex concepts into child-friendly educational comics, or automatically generate multimedia materials such as PPT, podcast scripts, and video content.
• data analysis : Automatically perform data-driven tasks such as data cleaning, visual chart generation, and business intelligence report writing. ©