Skip to content

DeerFlow

What it is

DeerFlow is an open-source agentic research workflow project from ByteDance focused on deep-research style information gathering and synthesis. By July 2026, it is recognized as a leading reference architecture for building high-autonomy research agents that utilize frontier models like Claude 4.8 Opus, GPT-5.5, and Gemma 3.

What problem it solves

It gives teams a structured, production-oriented starting point for building research agents instead of stitching together ad hoc search, scraping, and report-generation scripts. It addresses the complexity of multi-step browsing, information extraction, and the "hallucination-free" synthesis of large volumes of disparate data. It is particularly effective for automated benchmarking when combined with the MCP 3.0 Task Protocol.

Where it fits in the stack

Agents / Research Workflow. It sits between agent orchestration frameworks (like LangGraph) and end-user research products, providing a specialized layer for deep-search and synthesis loops. It leverages the MCP 3.0 ecosystem for standardized tool use and data retrieval.

Typical use cases

  • Strategic Intelligence: Compiling competitor, pricing, and tooling landscape reports.
  • Scientific Research: Gathering and summarizing academic papers and technical documentation.
  • Sales Enablement: Researching target accounts and public signals before outreach.
  • Content Creation: Building informed briefs and backgrounders for technical articles.
  • Automated Evaluation: Running research-heavy benchmarks as part of the MCP 3.0 Task Protocol.

Strengths

  • Reference Architecture: Provides a clear, battle-tested pattern for research-heavy workflows.
  • Open-Source: Highly adaptable and self-hostable, allowing for deep customization.
  • High Fidelity: Optimized for producing cited, evidence-backed reports.
  • Multi-Model Support: Native support for Claude 4.8 Opus, GPT-5.5, and Gemma 3.
  • Task Protocol Ready: Aligned with MCP 3.0 for structured research execution.

Limitations

  • Complexity: Requires significant adaptation for domain-specific production use cases.
  • Resource Intensive: Running deep-research loops can incur high token costs and require robust caching.
  • Evolving Ecosystem: Rapid changes in search APIs and model capabilities require frequent maintenance of the core workflow scripts.

When to use it

  • When you want a reference implementation for research-heavy agents that require evidence collection and synthesis.
  • When you are building a custom research assistant and need a head start on browsing and report generation logic.
  • When you need a self-hostable alternative to proprietary "AI Search" products.
  • For standardized research tasks that must comply with the MCP 3.0 Task Protocol.

When not to use it

  • For simple, single-step search tasks where a basic Tavily API call is sufficient.
  • When you require a stable, managed SaaS product with zero maintenance overhead.
  • For fast, transactional workflows that don't involve deep research or synthesis.

Getting started

Installation

The recommended way to start DeerFlow is via Docker to ensure all dependencies and execution environments are isolated: 

git clone https://github.com/bytedance/deer-flow.git
cd deer-flow
make config
make docker-init
make docker-start

Configuration

Update the generated config.yaml with your API keys for search providers and your preferred LLM (e.g., claude-4-8-opus-20260528 or gemma-3-27b).

CLI examples

# Initialize configuration and generate config.yaml
make config

# Start the application in development mode
make dev

# Run a specific research task via the CLI harness
python3 -m deerflow.harness run --task "competitor analysis for AI agents" --model "claude-4-8-opus"

API examples

import requests

# Example of submitting a research task to the DeerFlow API
url = "http://localhost:2026/api/v1/tasks"
payload = {
    "title": "Agentic Framework Research",
    "prompt": "Analyze the top 5 open-source agent frameworks in 2026.",
    "model_config": {
        "model": "claude-4-8-opus-20260528",
        "temperature": 0.0
    }
}
headers = {"Content-Type": "application/json"}

response = requests.post(url, json=payload, headers=headers)
print(response.json())

Sources / References

Contribution Metadata

  • Last reviewed: 2026-07-21
  • Confidence: high