DeerFlow Developer Guide
DeerFlow is an open-source deep-research framework built to combine planning, browsing, tool use, and human review into one repeatable workflow. This guide explains how DeerFlow is structured, how to get it running, and how to think about it as a developer.
This guide is based on the official DeerFlow docs and the official bytedance/deer-flow repository, checked on June 26, 2026.
In this section​
- Setup and first run
- Architecture and runtime
- Configuration and security
- Extending and contributing
- Simple user guide
1. What DeerFlow is​
The shortest accurate description is:
DeerFlow is a framework for running deeper, longer, more structured research workflows than a normal one-shot chat session.
The official docs position it around ideas like:
- planning before answering,
- browsing and evidence gathering,
- tool integration through MCP,
- human review in the loop,
- model-agnostic operation.
2. The mental model​
Think of DeerFlow as a research pipeline rather than as a simple chatbot:
| Layer | What it does |
|---|---|
| planner | Breaks a question into research tasks |
| runtime graph | Orchestrates multi-step execution |
| browser and tools | Gather evidence and external context |
| human review | Lets people steer or approve important steps |
| report output | Turns findings into usable artifacts |
3. Why developers would choose it​
- It is explicitly aimed at deep research, not casual prompting.
- It is model-agnostic.
- It supports MCP and human-in-the-loop workflows.
- It is built around a graph-based orchestration model, which makes it easier to reason about step boundaries.
4. When it makes sense​
DeerFlow is a good fit when your problem looks like:
gather evidence, compare sources, refine the plan, then write a result
It is less compelling when your problem is only "answer one short question quickly."
5. Recommended reading order​
- this overview,
- local install docs,
- harness and runtime docs,
- customization and MCP docs,
- the repo once the workflow shape is clear.