DeerFlow is an open-source super agent harness developed by ByteDance that orchestrates multiple components including sub-agents, memory systems, and sandboxes to handle complex, long-horizon tasks. The framework is designed to research, code, and create content, with capabilities spanning tasks that could take minutes to hours to complete. The project is written primarily in Python and maintains a homepage at deerflow.tech, positioning itself as a comprehensive agentic framework for handling diverse task complexity levels.
Version 2.0 represents a complete ground-up rewrite with no shared code from the original version, which is maintained separately on the 1.x branch. The framework integrates multiple architectural patterns including distributed task scheduling, workflow automation, event-driven processing, and asynchronous task execution. GitGenius classification data indicates the repository spans numerous technical domains including microservices orchestration, cloud-native applications, Kubernetes integration, and reactive programming patterns, reflecting its sophisticated infrastructure requirements.
The core feature set includes extensible skills and tools, sub-agent orchestration, sandbox environments for safe code execution, long-term memory capabilities, and context engineering. The framework supports Claude Code integration, session goals, and file system operations within sandboxes. Additional features include scheduled tasks, a terminal-based workbench interface, and embedded Python client support. The repository integrates with LangSmith and Langfuse for tracing, supports multiple instant messaging channels, and includes MCP server functionality. ByteDance has integrated InfoQuest, an intelligent search and crawling toolset, into DeerFlow for enhanced research capabilities.
Configuration is handled through an interactive setup wizard that guides users through LLM provider selection, optional web search configuration, and execution preferences including sandbox mode and bash access. The framework supports multiple model providers including Claude Code, Codex, OpenRouter, and vLLM deployments. Manual configuration is available through config.yaml with comprehensive examples for various provider setups. The project provides deployment sizing guidance ranging from local evaluation requiring 4 vCPU and 8 GB RAM to long-running servers needing 16 vCPU and 32 GB RAM.
Community engagement shows active development with 1579 tracked issues and pull requests. The median response latency is 2.7 hours with a mean of 202.5 hours, indicating responsive triage processes. The most active issue labels are bug (134 items), enhancement (90 items), and needs-triage (89 items). Primary contributors tracked by GitGenius include WillemJiang with 1546 events, hetaoBackend with 506 events, and foreleven with 346 events. The repository shares overlapping contributors with related projects including infiniflow/ragflow, langgenius/dify, and nousresearch/hermes-agent, suggesting an interconnected ecosystem of agentic frameworks.
The framework supports both Docker-based deployment with hot-reload capabilities for development and production configurations with local image builds. Linux with Docker is the recommended deployment target for persistent servers, while macOS and Windows are positioned as development environments. The project includes comprehensive tooling such as make doctor for setup verification and make support-bundle for diagnostic collection, facilitating issue reporting and troubleshooting. Security considerations are documented with recommendations for proper deployment to mitigate risks associated with sandbox access and file-write tools.
