DeerFlow

What is DeerFlow?

DeerFlow is an open source agent harness from ByteDance that runs long, multi-step research and creation tasks on your own infrastructure. A lead agent breaks a goal into sub-tasks, delegates them to specialized sub-agents (researcher, coder, reporter), and combines web search, code execution, and memory to hand back a finished deliverable — most often a cited report. It’s built on LangGraph and LangChain, with a Next.js web UI.

What is DeerFlow best for?

Developers and teams who want ChatGPT-style deep research without sending queries and documents to a vendor’s cloud, and who are comfortable running Docker. It fits people who need to bring their own models and search keys, inspect every step the agent takes, and extend the runtime with custom skills and tools rather than accept a fixed hosted product.

What can DeerFlow do?

• Run a supervisor + sub-agent workflow that plans a task and executes steps in sequence or parallel
• Search the web, crawl pages, and run Python in a sandboxed environment (local, Docker, or Kubernetes)
• Produce reports, slide decks, data analysis with visualizations, and podcast-style audio summaries
• Work with many model providers — OpenAI, Anthropic, Google Gemini, DeepSeek, Doubao, Qwen, and local vLLM or Ollama — with a different model per agent
• Keep long- and short-term memory across turns, plus checkpointing so you can branch or regenerate a conversation
• Extend the runtime with custom skills, Python functions, and MCP servers, and pipe results to Slack, Telegram, or Feishu

Where does DeerFlow fall short?

• It’s a fast-moving research project, not a stable product: the 2.0 release was a ground-up rewrite of the original “Deep Exploration and Efficient Research Flow” tool, so architecture and APIs shift between versions.
• The docs warn that improper deployment introduces real security risks — the agent runs generated code, so exposing it without proper authentication, HTTPS, and sandbox isolation is dangerous.
• It brings no model or search of its own. You supply and pay for LLM API keys (or run local models) and search providers, so “free” software still carries per-token inference costs.
• Multi-agent runs are resource-hungry — ByteDance suggests 8 vCPU / 16GB RAM for active use and 16 vCPU / 32GB for production workloads.

Is DeerFlow free?

Yes — DeerFlow is fully open source under the MIT license, free to self-host with no paid cloud tier or subscription. Your only costs are the server it runs on and the LLM and search API usage the agents consume (or the hardware to run local models).

What does DeerFlow replace?

DeerFlow is a self-hosted alternative to hosted deep-research and agent products: the deep research mode in ChatGPT, the AI answer engine Perplexity, and the general autonomous agent Manus. It does the same plan-search-synthesize job, but you run it yourself on your own models instead of paying a per-seat or credit-based subscription.

FAQ

Is DeerFlow open source? Yes. DeerFlow is released under the permissive MIT license, so you can self-host, modify, and use it commercially.

Can I self-host DeerFlow for free? Yes — the software is free and self-hosted only. You still pay for your own server and for the LLM and web-search API calls the agents make, unless you run local models.

Is DeerFlow a good Perplexity or Manus alternative? For technical users, yes — it gives you cited web research and autonomous task-completion on your own infrastructure and models. It asks more setup than a hosted app, so casual users wanting one-click answers may prefer the paid services.

What do I need to run DeerFlow? Docker, a machine with at least 4 vCPU and 8GB RAM (more for heavy use), and API keys for at least one LLM provider and a search provider. It ships a make setup wizard and Docker Compose to get started.