DeerFlow

Open-source multi-agent framework from ByteDance for long-running research, coding, and content tasks.

Free· Free, MIT-licensed (self-hosted; you pay only for LLM tokens)AgentsMulti-model (Doubao, DeepSeek, OpenAI, Gemini)

Best for

Pick DeerFlow if you want a self-hosted, MIT-licensed multi-agent harness for long-running research or coding tasks with full sandbox control.

Skip if

Skip it if you need a hosted, point-and-click agent product or you're not comfortable running Docker and managing your own LLM API keys.

DeerFlow is ByteDance's open-source 'SuperAgent harness' that orchestrates LLM-powered agents through long-horizon tasks — deep research, code generation, multimedia creation — that can run for minutes or hours. The framework ships with a persistent Docker sandbox (file system, shell, VSCode integration), long- and short-term memory, sub-agent spawning, progressive skill loading, and a planner capable of decomposing complex problems into sequential or parallel sub-tasks.

It's pitched at developers and researchers who want an autonomous agent stack they can self-host and fully control rather than a closed SaaS like Manus or Devin. DeerFlow is MIT-licensed (the repo lives at bytedance/deer-flow) and brings its own model-agnostic routing: out of the box it speaks to Doubao, DeepSeek, OpenAI, and Gemini, so you can mix a cheap planner with a stronger executor. Because it's OSS and self-hosted, there's no per-seat or per-task pricing — your only cost is the LLM tokens and the host you run the sandbox on.

The trade-off is the usual one for agent frameworks at this layer: you need Docker, an LLM key, and the patience to wire up your own tools and prompts. There's no managed SaaS, no hosted web app for non-technical users, and the docs/community are still maturing compared with LangGraph or CrewAI.

Editor's take

A serious open-source entrant from ByteDance that sits between LangGraph's plumbing and Manus's closed UX — you get the persistent sandbox and sub-agent loop without giving up control of your stack. Worth a look if you're building agentic research workflows and have rejected SaaS lock-in, but expect to do real engineering.

— The AI Tool Bible editorial team