ByteDance just open-sourced DeerFlow, a local-first “super agent” harness that can research, code, build sites, and create decks on your machine.DeerFlow (Deep Exploration and Efficient Research Flow) is ByteDance’s MIT-licensed agent framework. Version 2 is a full rewrite built around a harness that coordinates sub-agents, persistent memory, sandboxed execution, and extensible skills. It is designed for longer tasks that can run from minutes to hours, not single chat replies.What it can do, according to the project:
- Deep research and report writing
- Code generation and execution inside an isolated sandbox (local or Docker)
- Website and page building
- Slide-deck creation
- Image and video generation through skills
- File-system and terminal access so the agent can actually run commands and write files, not only suggest them
It is fully open source, runs on your own hardware via Docker or a local setup, and does not require a DeerFlow subscription. You still need an LLM: it works with OpenAI-compatible APIs, models such as DeepSeek or Claude, or local models through Ollama. Data stays on your machine only if you also keep the model local; cloud model calls send prompts off-device.Setup is a clone plus configuration:
bash
git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make setupIt is an agent orchestrator, not a finished “employee” that replaces a full-time role out of the box. Results depend on the model, skills, and how carefully you scope the task. For teams that want a self-hosted alternative to closed agent platforms, the repo is the place to
Project Link to start: https://github.com/bytedance/deer-flow