Overview

DeerFlow 2.0 (Deep Exploration and Efficient Research Flow) is ByteDance's open-source SuperAgent harness — a complete ground-up rewrite that evolved from an internal deep-research tool into a full-featured, batteries-included agent runtime. It orchestrates a lead agent with specialized sub-agents, persistent memory, secure Docker sandboxes, and extensible skills to tackle sophisticated, long-running tasks that can span minutes to hours.

Built on LangGraph and LangChain, it ships with everything needed out of the box: a virtual computer (filesystem + bash), context engineering, and model-agnostic LLM support. It reached #1 on GitHub Trending shortly after launch in February 2026.

Key Features

Multi-Agent Orchestration: Lead agent plans tasks and spawns parallel sub-agents with isolated contexts, tools, and termination conditions.
Secure Sandbox Environment: Docker-based (AIO Sandbox) with persistent, mountable filesystem, bash terminal, code execution, and browser support — isolated, safe, and long-running.
Progressive Skills System: On-demand loading of modular skills (deep-search, report generation, frontend-design, deploy, etc.) to keep context windows efficient; fully extensible with custom SKILL.md files.
Advanced Memory: Short-term summarization + long-term persistent memory across sessions (user preferences, knowledge, and workflows).
Context Engineering: Aggressive summarization, offloading to filesystem, and isolated sub-agent contexts for handling very long tasks.
Model Agnostic: Works with any OpenAI-compatible API (GPT, Claude, Gemini, DeepSeek, Doubao, Kimi, etc.).
Integrations: Claude Code terminal support, IM channels (Telegram, Slack, Feishu/Lark), MCP servers, and Python client.

How It Works

Ingest tasks via web UI, Python client, or messaging bots.
Lead agent decomposes the goal and spawns sub-agents.
Sub-agents use relevant skills and sandbox execution.
Results are synthesized, summarized, and stored in memory/filesystem.
Final deliverable (reports, code, websites, slides, videos) is produced.

Example Python client usage:

from deerflow.client import DeerFlowClient
client = DeerFlowClient()
response = client.chat("Analyze this dataset and build a dashboard", thread_id="analysis-1")

Quickstart

git clone https://github.com/bytedance/deer-flow.git && cd deer-flow
make config
make docker-init
make docker-start

Access at http://localhost:2026. Configure models and API keys in config.yaml and .env.

Use Cases

Deep research and comprehensive report generation
Full-stack coding projects with file I/O and deployment
Exploratory data analysis and visualizations
Content creation (slide decks, websites, video generation)
Multi-step automation workflows across tools and data sources
Enterprise knowledge work and internal tooling

License and Availability

Fully open source under the MIT license. Self-hostable with Docker or local development. Contributions welcome.

For demos, skills library, and latest updates visit the official site or GitHub repository.