What it does
DeerFlow 2.0 is a ground-up rewrite of an earlier deep-research framework, which is still maintained on the 1.x branch. It calls itself a super agent harness that orchestrates sub-agents, memory and sandboxes, powered by extensible skills, for tasks that can take minutes to hours. Setup is clone-and-run. A wizard asks for an LLM provider, an optional web search provider and safety preferences such as sandbox mode, bash access and file-write tools, then writes the config. Docker is the recommended route, local development is also supported, and make doctor checks the installation. Models are set up per provider with an API key and model ID. Presets cover OpenAI, Anthropic, DeepSeek, Qwen, Doubao, Kimi, Zhipu and MiniMax, and custom OpenAI-compatible endpoints, vLLM and Ollama are supported. Web search and page-fetch tools work with several providers. The documentation outline also lists an MCP server, IM channels, tracing with LangSmith, Langfuse or Monocle, scheduled tasks, projects, long-term memory, an embedded Python client and a terminal workbench. The page carries a security notice warning that improper deployment can introduce risks.
Capabilities
- Sub-agents, memory and sandboxes working under a lead agent
- Extensible skills and tools, with an MCP server
- Setup wizard and Docker deployment
- Many model providers and custom endpoints
- Scheduled tasks, projects and long-term memory
- Tracing with LangSmith, Langfuse or Monocle
When to use it
- You want long, multi-step research or coding tasks run by sub-agents on your own infrastructure
- You want to plug in your own model provider and web search tools
- You are building agent features and want a harness with sandboxes and memory to start from
