DeerFlow

DeerFlow is an open-source SuperAgent harness that autonomously researches, codes, and creates by combining sandboxes, memories, tools, skills, and subagents into a single platform for tasks ranging from minutes to hours.

What is DeerFlow?

DeerFlow (https://deerflow.run/) is a free, open-source (MIT License) AI agent harness built by ByteDance that takes high-level user intents—such as "Research mRNA delivery, build a landing page, deploy to Vercel"—and autonomously executes them using sandboxed code execution, long/short-term memory, and multi-model support (Doubao, DeepSeek, OpenAI, Gemini). It produces finished outputs including code, webpages, videos, research reports, and comic strips.

Key Features

• All-in-One Sandbox — Combines Browser, Shell, File, MCP, and VSCode Server in a single Docker container for isolated, persistent, and long-running agent execution.
• Long/Short-term Memory — Retains context across sessions so the agent adapts to individual user workflows over time.
• Planning and Sub-tasking — Breaks complex objectives into sequential or parallel subtasks, replans on failure, and executes autonomously.
• Skills and Tools — Extensible dynamic library invoked via .skill files; users can author and share custom skills (example skills on the site include deep-search with subject bundles like biotech, computer science, physics, and frontend-design with an accompanying deploy skill).
• Multi-Model Flexibility — Supports Doubao, DeepSeek, OpenAI, and Gemini, letting users choose the model for each task.
• Progressive Skill Loading — Skills are loaded on demand to minimize overhead, with the ecosystem open to third-party .skill files for bespoke workflows.

Who is it for?

• Developers and researchers who need an autonomous agent to run complex multi-step projects (e.g., data analysis on the Titanic dataset with visualizations, deploying a landing page to Vercel).
• Content creators and educators who want to generate videos, comic strips, or research reports from novel scenes or technical concepts (e.g., Doraemon explaining MOE architecture).
• Technical startup founders and analysts who use the agent to watch Y Combinator videos, synthesize insights, or forecast industry trends.

What can you do with DeerFlow?

• Deep research and forecasting — Generate a webpage with a deep research report forecasting agent technology trends and opportunities in 2026.
• Creative synthesis — Search a scene from "Pride and Prejudice", then generate a video and a reference image based on the scene.
• Educational content — Create a Doraemon comic strip that explains Mixture of Experts (MOE) architecture to AI-interested teenagers.
• Podcast analysis — Collect all podcast appearances of Dr. Fei-Fei Li in the last 6 months and summarize them into a comprehensive report.

How does DeerFlow work?

Users provide a natural-language goal in the workspace, and DeerFlow plans the task, delegates subtasks to specialized subagents, executes code and tool calls inside a sandboxed environment, and returns the final output. The agent can browse the web, run shell commands, read/write files, and use MCP servers—all within a single Docker container.

Pricing

DeerFlow is free and open source (MIT License). Users can self-host the entire stack for full control.

FAQ

Is DeerFlow free to use?

Yes. DeerFlow is released under the MIT License, meaning it is free to use, modify, and self-host without any licensing fees.

What models does DeerFlow support?

It supports Doubao, DeepSeek, OpenAI (including GPT-4 series), and Gemini. Users can switch models per task.

Can I add my own skills?

Yes. DeerFlow uses .skill files that define custom capabilities. You can write your own skills and load them into the agent, expanding its functional reach.