DeerFlow
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.
Last verified:
What is DeerFlow?
DeerFlow is an open-source SuperAgent harness developed by ByteDance for building and operating autonomous agent systems. It provides a runtime framework that orchestrates agents with memory, tools, skills, sandboxes, and sub-agents to handle complex, long-horizon tasks that can take minutes to hours. The framework combines large language models with automated tool-use capabilities including web search, web crawling, Python execution, and data processing to enable end-to-end research workflows.
The platform features two main components: the DeerFlow Harness (core SDK and runtime layer for building custom agent systems) and the DeerFlow App (reference application for deployment and end-user workflows). Key capabilities include multi-agent architecture with specialized roles (Coordinator, Planner, Researcher, Coder, Reporter), Docker-based isolated sandboxes for secure code execution, short-term and long-term memory systems, extensible skills defined as Markdown files, and the ability to produce multimodal outputs including reports, PowerPoint presentations, code, data visualizations, and AI-generated podcast audio.
DeerFlow is designed for developers, research teams, software engineers, content creators, and enterprises who want to build real autonomous agents with full data sovereignty. It suits use cases like deep research and literature reviews, competitive intelligence analysis, automated content marketing pipelines, code generation and testing, data analysis with visualizations, website creation, and presentation generation. The framework is built on LangGraph and LangChain, supports 100+ models through litellm, and integrates with MCP servers for tool invocation.
DeerFlow pricing
Pricing model: Freemium
DeerFlow is fully free and released under MIT license allowing unrestricted commercial use. There are no subscription plans or paid tiers. The only costs users incur are their own infrastructure expenses (VPS, cluster, or private cloud hosting) and LLM API usage fees. This open-source structure often beats SaaS pricing for teams with operational skills. The project is self-hosted, meaning you run it on your own servers or local machines.
DeerFlow pros
- Fully open-source MIT license with unrestricted commercial use
- Self-hostable with complete control over data and models
- Multi-agent architecture with 5 specialized agent roles
- Docker-based isolated sandboxes for secure code execution
- Persistent filesystem for file writing and long-running tasks
- Both short-term and long-term memory systems
- Extensible skills system using Markdown files
- Supports 100+ LLM providers through litellm
- Produces multimodal outputs: reports, PPT, code, podcasts, videos
- Sub-agent spawning with parallel execution capability
- Asynchronous task coordination for complex workflows
- Human-in-the-loop mechanism for live plan revision
- Built-in RAG knowledge base retrieval integration
- MCP protocol support for private knowledge onboarding
- Context engineering with automatic task summarization
- Integration with LangSmith and Langfuse for tracing
- Messaging channel support for Slack and Telegram
- Ground-up rewrite in v2.0 with no legacy code burden
DeerFlow cons
- Requires technical skills to deploy and operate
- Only costs are infrastructure and LLM API usage (no free SaaS)
- Complex compared to simple AI chat applications
- Security risks if improperly deployed beyond trusted environments
- Requires capable underlying models with long context
- Linux and Docker preferred for persistent deployment
- Teams need engineering judgment and security review
- Not a closed black-box SaaS solution
Frequently asked questions about DeerFlow
What is DeerFlow 2.0?
DeerFlow 2.0 is ByteDance's open-source super agent harness for long-horizon AI work. It is a ground-up rewrite that shares no code with v1, positioning itself as a full super agent harness rather than just a deep research framework. It orchestrates sub-agents, memory, sandboxes, files, and extensible skills to handle research, coding, asset creation, and multi-step tasks taking minutes to hours.
Is DeerFlow open source?
Yes, DeerFlow 2.0 is MIT licensed and available as a self-hosted open-source project on GitHub. The code is fully inspectable, usable, and modifiable by the community with unrestricted commercial use allowed.
Does DeerFlow support only one model provider?
No, DeerFlow is designed around OpenAI-compatible APIs and supports 100+ models through litellm. It works with multiple providers including Doubao, DeepSeek, OpenAI, Gemini, Claude, Qwen, and Kimi, improving switching flexibility and cost control.
What is DeerFlow best at?
DeerFlow excels at multi-step, tool-heavy, output-driven tasks. It is strong for deep research with structured output, competitive intelligence analysis, academic literature reviews, automated content marketing pipelines, code generation with test execution, data analysis with visualizations, website creation, presentation generation, and multimodal content producing reports, slides, and podcasts from single prompts.
How does DeerFlow handle security?
DeerFlow uses Docker-based isolated sandboxes for secure code execution, preventing agents from accessing the host system directly. The project recommends strict security measures including sandbox mode, careful networking decisions, and access control. ByteDance explicitly warns that improper deployment beyond trusted local environments can introduce serious security risks since the system can execute commands and manage files.
What are skills in DeerFlow?
Skills are structured Markdown capability modules that define workflows, best practices, and supporting resources for specific task classes. They are loaded progressively on demand to keep context windows leaner, and developers can add or replace skills as workflows change without touching the core framework. Built-in skills include research, report generation, slide creation, web pages, image generation, and video generation.
How do sub-agents work in DeerFlow?
DeerFlow can spawn sub-agents with scoped context, tools, and termination conditions. These sub-agents handle sub-tasks and can run in parallel when needed, returning structured results to the lead agent which then synthesizes the final output. This enables task decomposition for long-horizon work where the coordinator agent breaks missions down and delegates to specialized agents.
What is the difference between DeerFlow Harness and DeerFlow App?
DeerFlow Harness is the core SDK and runtime layer for building your own custom agent system, written for developers who want to integrate DeerFlow capabilities or build custom agent products. DeerFlow App is a reference application built on top of the Harness for teams who want to deploy DeerFlow as a complete usable product with configuration, operations, and end-user workflows.
How do I deploy DeerFlow?
The official quick start involves cloning the GitHub repo, running 'make setup', choosing an LLM provider and safety preferences, then starting through Docker or local development. Docker is the recommended path for persistent deployment, with the app accessible at http://localhost:2026 by default. Linux plus Docker is the preferred target for production deployment.
What outputs can DeerFlow produce?
DeerFlow produces multimodal outputs beyond plain text, including structured research reports, PowerPoint slide decks, web pages, code with test suites, data visualizations and charts, AI-generated podcast audio through TTS, images, videos, and comic strips. This goes beyond text-only research tools to create actual deliverables from single natural language prompts.