DeerFlow

DeerFlow is a self-hosted AI agent application for research, coding and content creation on your own infrastructure. It supports local models through Ollama and vLLM, including Qwen3, as well as hosted providers. With a cloud provider, model requests leave your server and may incur provider charges.

Agents can execute commands, manage files and work on tasks that take minutes or hours. Docker sandboxes provide an isolated workspace; the recommended All-in-One Sandbox includes a browser, shell, file tools and VSCode Server. Local host execution and Kubernetes-based sandboxes are also supported.

Built-in skills cover research reports and slides, alongside webpages, image and video generation. DeerFlow can analyze datasets and produce visualizations. You can add or replace skills, combine workflows and transfer custom skill packages between instances. It loads skills as needed rather than putting the whole collection into each model request.

Optional RAGFlow retrieval connects agents to private knowledge datasets. Clickable citations show retrieved excerpts and document names, plus page numbers when RAGFlow supplies them. Evidence snapshots stay attached to conversations, but standalone Markdown and batch exports don't carry those interactive records.

DeerFlow is open source under MIT. Linux with Docker is the recommended server environment; macOS and Windows support development and evaluation. A light local evaluation starts around 8 GB RAM, with local model hardware needs assessed separately. MCP tools and Lark/Feishu integrations extend its connections.