DeerFlow

DeerFlow is an MIT-licensed AI agent framework for people who want to build and run agents on their own infrastructure. It combines subagents, memory and a Docker-based workspace to handle research, coding and content creation tasks that can take minutes or hours.

Its agents can execute commands, work with files and carry out longer jobs inside a sandbox. The recommended All-in-One Sandbox brings a browser, shell, file tools, MCP and VSCode Server into one container. This gives agents a working environment for producing outputs such as webpages, reports and data visualizations.

Skills define the work. Built-in skills cover research, slide creation, image and video generation, alongside other document and web tasks. You can add your own workflows, replace existing skills or combine them. DeerFlow loads skills only when a task needs them, which reduces the amount of context they consume. Custom skills can also move between DeerFlow instances as exportable archives.

For work involving private documents, DeerFlow can connect to RAGFlow and retrieve evidence from permitted datasets. Answers can include clickable citations with source excerpts, document names and page numbers when available. Those evidence snapshots stay with the conversation and remain inspectable after reloading it.

DeerFlow uses LangChain and supports MCP tools and Lark/Feishu integration. Its Python framework is open source under the MIT license.