DeerFlow

DeerFlow (Deep Exploration and Efficient Research Flow) is an open-source SuperAgent harness built by ByteDance under the MIT License. It orchestrates sub-agents, memory, sandboxed code execution, and extensible skills to tackle long-horizon tasks that range from deep research and report generation to coding, video creation, and data analysis.

Built on LangChain, DeerFlow is designed to handle tasks that take minutes to hours by decomposing complex objectives into manageable sub-tasks executed by specialized sub-agents. Each sub-agent operates in its own isolated context with dedicated tools, sandboxed execution environments, and structured termination conditions, while the lead agent synthesizes results into coherent deliverables.

With its 2.0 release, DeerFlow evolved from a Deep Research agent into a full-stack Super Agent, adding long/short-term memory, planning and sub-tasking, an extensible skills system, multi-model support, and a persistent sandbox with a mountable file system. It reached the #1 spot on GitHub Trending shortly after launch and has garnered over 76,000 stars on GitHub.

• Deep Research - Browses the web, collects data from multiple sources, and generates comprehensive research reports. Supports multi-angle exploration with parallel sub-agents for thorough coverage of complex topics.
• Code Generation & Execution - Writes, edits, and executes code inside secure sandboxed containers. Supports multiple programming languages with full file system access and package installation.
• Video & Image Generation - Creates text-to-video, image-to-video, and static visual content. Can research source material, generate reference images, and produce video output from textual or visual prompts.
• Multi-Model Support - Works with Doubao, DeepSeek, OpenAI, Google Gemini, and any OpenAI-compatible API. Swap models without restarting and use different models for different sub-tasks.
• Sandboxed Execution - Each task runs in an isolated Docker-based sandbox with its own file system, shell, and browser. The AIO Sandbox combines browser, shell, file operations, MCP, and VS Code Server in a single container.
• Persistent Memory - Remembers user preferences, writing style, technical stacks, and recurring workflows across sessions. Memory is stored locally and stays under user control.
• Extensible Skills System - Skills are Markdown files that define workflows, best practices, and resource references. Loaded progressively only when needed. Add custom skills or use built-in ones for research, reports, slides, web pages, and more.
• Agent Sub-Delegation - Spawn isolated sub-agents with scoped context, tools, and termination conditions. Sub-agents run in parallel when possible and report structured results back to the lead agent.
• Visual Canvas Interface - A Web UI workspace with a chat interface, file browser, workspace editor, and active goal display offering a visual overview of ongoing tasks and outputs.
• Scheduled Tasks - Built-in scheduled task system for automating recurring research, report generation, and monitoring workflows without manual intervention.

DeerFlow is built on LangChain and follows a modular agent architecture designed for long-horizon task execution:

• LangChain Foundation - The core orchestration layer uses LangChain for LLM integration, tool calling, and conversation management. It provides a unified interface across multiple model providers with support for function calling and structured outputs.
• Lead Agent with Sub-Agent Delegation - A lead agent receives the user's objective and decomposes it into sub-tasks. Sub-agents are spawned with isolated context windows, dedicated tool sets, and clear termination conditions. They run sequentially or in parallel, returning structured results for synthesis.
• Sandboxed Execution Environment - Each task gets its own Docker-based sandbox (AIO Sandbox recommended) with a full file system at /mnt/user-data/ containing uploads, workspace, and outputs directories. The sandbox supports shell commands, file operations, browser automation, and MCP servers.
• Skills System - Skills are Markdown capability modules loaded progressively on demand. Built-in skills include research, report generation, slide creation, web page development, and image/video generation. Custom skills can be added by placing SKILL.md files in the skills directory.
• Context Engineering - Each sub-agent runs in an isolated context to stay focused. DeerFlow manages context aggressively through summarization, offloading intermediate results to the file system, and compressing irrelevant history.
• Long-Term Memory - Across sessions, DeerFlow builds a persistent memory of user profiles, preferences, and knowledge. Memory is stored locally and deduplicates entries to avoid accumulating repeated facts.
• Session Goals - Users can set active completion conditions with /goal. The agent evaluates progress against the goal after each run, automatically continuing until the condition is met or a safety cap is reached.

DeerFlow is model-agnostic and works with any LLM that implements the OpenAI-compatible API. The model system is flexible enough to support different providers for different use cases, and you can configure multiple models in a single configuration file.

Recommended models for DeerFlow include those with long context windows (100k+ tokens), strong reasoning capabilities, multimodal inputs, and reliable tool-use for function calling.

Supported Models

Provider	Model	Notes
Doubao	Doubao-Seed-2.0-Code	ByteDance's own model, strongly recommended for coding tasks
DeepSeek	DeepSeek V3.2	Excellent reasoning and long-context capabilities
OpenAI	GPT-4o, GPT-5	Full support including Responses API
Google	Gemini 2.5 Flash/Pro	Via OpenRouter or direct API
Kimi	Kimi 2.5	Recommended for general reasoning tasks
OpenRouter	Various	Gateway to 200+ models via unified API
Local / vLLM	Qwen3, Llama, etc.	Self-hosted via vLLM or Ollama
CLI-backed	Claude Code, Codex CLI	Provider integrations via CLI OAuth

Configuration is done through a single config.yaml file where each model entry specifies the provider class, model name, API key, and optional parameters like base URL, reasoning effort, or max tokens. Models can be swapped dynamically without restarting the server.

• Deep Research Reports - Deploy sub-agents to explore multiple angles of a topic simultaneously. Collect data from web sources, YouTube videos, and podcasts, then synthesize findings into comprehensive, well-structured reports with citations.
• Code Generation & App Development - Generate complete web applications, scripts, and tools inside the sandbox environment. DeerFlow can create wireframes, write code, install dependencies, test, and deploy in a single workflow.
• Video & Creative Content Creation - Generate text-to-video and image-to-video content for storytelling, marketing, or educational purposes. Research source material, create reference images, and produce video output from textual prompts.
• Data Analysis & Visualization - Upload datasets and let DeerFlow perform exploratory data analysis, generate visualizations, identify trends, and produce insight reports with charts and statistical summaries.
• Slide Deck & Web Page Generation - Create presentation slides or complete web pages from natural language descriptions. DeerFlow researches content, designs layouts, and generates output files ready for presentation or deployment.
• Multi-Agent Collaboration - Use the visual canvas to orchestrate multiple agents working on different aspects of a project. Each agent operates in its own context while contributing to a shared workspace.

The fastest way to try DeerFlow is through the official website at deerflow.tech. The website features an interactive workspace with case studies, demos, and a visual canvas where you can experience DeerFlow's capabilities directly.

One-Line Agent Setup:

Help me clone DeerFlow if needed, then bootstrap it for local development by following
https://raw.githubusercontent.com/bytedance/deer-flow/main/Install.md

Self-Hosted Deployment

For those who want to self-host, DeerFlow is available on GitHub under the MIT License. Docker is the recommended deployment method for production use:

git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make setup            # Interactive configuration wizard
make docker-init      # Pull sandbox image
make docker-start     # Start services

Deployment Sizing

Target	Minimum	Recommended
Local development	4 vCPU, 8 GB RAM	8 vCPU, 16 GB RAM
Docker development	4 vCPU, 8 GB RAM	8 vCPU, 16 GB RAM
Long-running server	8 vCPU, 16 GB RAM	16 vCPU, 32 GB RAM

For production, Linux plus Docker is the recommended deployment target. macOS and Windows are best treated as development or evaluation environments. Access the Web UI at http://localhost:2026 after starting the services.

Detailed documentation is available on the official website. The GitHub repository includes a complete config.example.yaml reference, Docker production deployment commands (make up / make down), and a local development workflow (make dev) for contributors.