deer-flow
An open-source long-horizon SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skill, subagents and message gateway, it handles different levels of ta
About
An open-source long-horizon SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skill, subagents and message gateway, it handles different levels of tasks that could take minutes to hours.
Details
DeerFlow is an open-source SuperAgent harness designed for long-horizon tasks that involve deep research, coding, and content creation. It utilizes sandboxes, memories, tools, skills, subagents, and a message gateway to manage complex workflows spanning minutes to hours. The framework operates in a secure Docker-based runtime environment, including an All-in-One Sandbox with browser, shell, file management, VSCode server, and persistent file systems, enabling safe execution of commands, long-running processes, and file persistence. Users benefit from its extensibility, allowing custom skill files or built-in libraries, and multi-model support for providers like Doubao, DeepSeek, OpenAI, and Gemini.
DeerFlow 2.0 enhances capabilities with context engineering via long- and short-term memory, advanced planning and sub-tasking for sequential or parallel execution, and plug-and-play tools. Case studies illustrate its versatility, such as generating webpages with research forecasts, videos and images from novels like Pride and Prejudice, comic strips explaining AI concepts like MOE architecture, exploratory data analysis on datasets like Titanic with visualizations, deep research from YouTube videos, and podcast summarization. This makes it suitable for researchers, developers, content creators, and data analysts seeking a self-hosted, controllable agent platform.
As a fully open-source project under MIT license, DeerFlow emphasizes community collaboration, with GitHub contributions welcomed for shaping its evolution from a deep research agent to a full-stack SuperAgent.
Key Features
- Docker-based secure sandbox with browser, shell, file management, and VSCode
- Long- and short-term memory for context engineering
- Extensible skills and tools with progressive loading
- Planning and sub-tasking for long-horizon tasks
- Subagents and message gateway for task handling
- Multi-model support including Doubao, DeepSeek, OpenAI, Gemini
- Persistent file system for read/write operations
- Open-source MIT license for self-hosting
Use Cases
- Deep research and report generation with webpage output
- Video and image generation from literary scenes or novels
- Comic strip creation for explaining technical concepts
- Exploratory data analysis with visualizations on datasets
- Analysis of videos followed by deep research reports
- Collection and summarization of podcasts into reports
Pros & Cons
Pros
- Fully open-source under MIT license for self-hosting and full control
- Handles long-horizon tasks with planning, subagents, and memory
- Secure, isolated sandbox environment for safe execution
- Flexible multi-model support for various LLM providers
- Extensible with custom skills, tools, and community contributions
- Appears free with no mentioned usage limits beyond self-hosting resources
Cons
- Requires technical setup for Docker-based self-hosting and sandbox management
- Resource-intensive for long-running tasks; hardware limits should be considered
- Output quality varies by underlying models, prompts, and skills used
- May need internet access for certain tools, models, or data sources
- As open-source, lacks official support; relies on community for troubleshooting
Frequently Asked Questions
Is DeerFlow free to use?
It appears to be fully free and open-source under MIT license; self-hosting costs depend on infrastructure
What models does DeerFlow support?
Based on available information, it supports Doubao, DeepSeek, OpenAI, Gemini, and others; full list should be verified in documentation
Does it handle multimedia outputs like video and images?
Case studies show video generation, image creation, and comic strips; capabilities depend on integrated tools and models
Is a sandbox environment provided?
Yes, it includes a Docker-based All-in-One Sandbox with browser, shell, VSCode, and persistent FS; setup required
Can users extend DeerFlow with custom features?
Yes, it supports extensible skills, tools, and subagents; built-in library and custom files available
What license does it use?
MIT License, allowing full control and modifications; confirmed on the website