DeerFlow
An open-source agent framework that coordinates sub-agents, memory, and sandboxes for long, multi-step tasks.
- Stars
- 84.9k
- Forks
- 11.9k
- Licence
- MIT
- Language
- Python
- Last update
- 4 days ago
- Type
- AI model / agent
What it is
DeerFlow is an open-source 'super agent harness' that runs an AI lead agent which can split work across sub-agents, keep long-term memory, and execute code or tasks inside sandboxes. It is built to handle tasks that may take minutes to hours, such as research, coding, and content creation. The README describes it as extensible through skills and tools.
- Run a multi-step research task where the agent gathers sources and writes a summary
- Delegate coding subtasks to sub-agents inside an isolated sandbox
- Connect the agent to external chat channels via the IM Channels feature
- Query private documents through the RAGFlow-based knowledge retrieval feature
Great for
- Developers who want to experiment with multi-agent workflows on their own infrastructure
- Researchers running long, multi-step investigations that need memory and sub-agents
- Teams already using LangGraph or LangChain who want a reference agent harness
- Users who are comfortable configuring their own LLM API keys and running services locally or in Docker
Not ideal for
- Non-technical users looking for a ready-to-use consumer app, since setup involves configuration, Docker or local development, and API keys
- Anyone who needs a fully managed, hosted service with no operational work
- Security-sensitive deployments without reviewing the README's security notice, because the gateway admin role is equivalent to code execution
How to get it
- 1
Run with Docker
You'll need: Python 3.12+ · Node.js 22+ · Docker (recommended per README)
Clone the repository
git clone https://github.com/bytedance/deer-flow.git - 2
Run it
- Clone the repository and review the Configuration section of the README
- Set up your model provider configuration as described in the docs
- Follow the Docker option (recommended in the README) to start the application
- Open the local web interface once the services are running and try a task
- 3
Get the most out of it
- Review the Recommended Models section to choose a model suited to the harness
- Enable Sandbox Mode to run agent code in isolation
- Explore the MCP Server and Skills & Tools sections to extend capabilities
- Use the Long-Term Memory and Session Goals features for multi-session work
- Read the Security Notice before exposing any deployment to others
Downloads
Strengths
- Open-source under the MIT license, so the code can be inspected and modified
- Combines sub-agents, memory, sandboxes, and skills in one harness
- Documented deployment options including Docker and local development
Trade-offs
- Setup requires multiple prerequisites and configuration, which the README describes as non-trivial
- The README includes a security notice warning that improper deployment can introduce risks
- The README is long and truncated here, so some details of setup and behavior could not be verified
How it compares
| Project | Stars | Type | Level | Updated | Best for |
|---|---|---|---|---|---|
| deer-flow (this) | 84.9k | AI model / agent | advanced | 4 days ago | Developers who want to experiment with multi-agent workflows on their own infrastructure |
| langgraph | 44.3k | Library | intermediate | 4 days ago | Python developers building multi-step AI agents that must survive failures or long waits |
| crewAI | 60.4k | Library | intermediate | 4 days ago | Python developers prototyping multi-agent workflows |
| OpenHands | 92.5k | App | intermediate | 4 days ago | Developers who want coding agents running on a server while their laptop is off |
| ruflo | 80.6k | Dev tool | intermediate | 4 days ago | Developers already using Claude Code or Codex who want multiple agents working together |
| learn-harness-engineering | 22.9k | Learning resource | beginner | 14 days ago | Developers new to building or controlling AI coding agents |
Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.