DeerFlow

1. Project Overview

The full name of the DeerFlow is Deep Exploration and Efficient Research Flow. It was originally a Deep Research framework, and then the community extended it to more scenarios such as data pipeline, presentation generation, Dashboard, content automation, etc., so it was 2.0 refactored into a more general Super Agent Harness.

Pre-sales can be interpreted as an open source Agent platform base that equips large models with "working environment, long-term memory, skill system, file system, tools and sub-agents.

2. What does it mostly do?

2.1 in-depth research and report generation

DeerFlow inherits the core capabilities of Deep Research. It can search for data, crawl web pages, organize information, analyze from multiple perspectives, and output reports around a complex problem.

Suitable scenarios:

2.2 long task execution

The website emphasizes that DeerFlow can handle tasks ranging from a few minutes to a few hours. It supports long-link tasks through planning, subtask disassembly, file systems, memory, and sandbox execution environments.

Typical tasks:

-Research an open source project and generate an analysis report.

-Read uploaded data and generate PPT or web pages.

-Write a small application or script.

-Analyze data files and produce charts and interpretations.

-Multiple rounds of iteration over a content item or technical solution.

2.3 Sub-agents Parallel Collaboration

The DeerFlow lead agent can dynamically pull up sub-agents. Each sub-agent has an independent context, tool, and termination condition, can process different directions in parallel, and is finally aggregated by the lead agent.

Presentation:

DeerFlow is not just a single agent in a context, but can split complex tasks into multiple sub-tasks, allowing different sub-agents to study, execute, and summarize in parallel, which is more suitable for complex business scenarios.

2.4 Sandbox and File Systems

DeerFlow provide an executable environment for the Agent. Each task has isolated working directories, including upload files, workspace, and output directories:

/mnt/user-data/
├── uploads/          ← 用户上传文件
├── workspace/        ← Agent 工作目录
└── outputs/          ← 最终交付物

It supports:

-Read and write files

-Edit files

-Execute bash or code

-View pictures

-Generate final deliverables

-Isolate execution environment in Docker or Kubernetes dependent mode

This brings it closer from "a chatbot that invokes tools" to "an agent with a controlled computer".

2.5 Skills and Tool Extensions

DeerFlow supports the standard Agent Skill. A Skill is typically a 'SKILL.md' that can contain workflows, best practices, reference materials, scripts, and templates.

Built-in and adaptable to include:

-research

-report-generation

-slide-creation

-web-page

-image-generation

-video-generation

-claude-to-deerflow

Example structure:

/mnt/skills/public
├── research/SKILL.md
├── report-generation/SKILL.md
├── slide-creation/SKILL.md
├── web-page/SKILL.md
└── image-generation/SKILL.md

/mnt/skills/custom
└── your-custom-skill/SKILL.md

Core value: Enterprises can solidify internal methodology, industry knowledge, delivery templates, and pre-sales processes into Skill to form a reusable Agent capability library.

2.6 long-term memory

DeerFlow supports cross-session memory to gradually accumulate user portraits, preferences, background knowledge, work habits, and common processes.

Suitable for:

-Personal work assistant

-Team knowledge assistant

-Long-term customer follow-up assistant

-Fixed business process automation assistant

2.7 multi-end entrance

DeerFlow is not just a backend SDK, it provides multiple entries:

3. Project comes with schematics and official visuals

3.1 official website positioning map and demonstration entrance

The official website describes DeerFlow as an open source SuperAgent Harness that can be researched, coded and authored. With sandbox, memory, tools, skills and subagents, long tasks can be handled at different levels.

-Official website:https://deerflow.tech

-README also contains a link to the official demo video:deer-flow-720p.mp4

3.2 InfoQuest Integration Chart

BytePlus InfoQuest search and crawler integration is demonstrated in DeerFlow README.

! InfoQuest Banner

Pre-sales implications: DeerFlow not only rely on the model itself in the "deep research" scenario, but also emphasize search, crawling and data acquisition capabilities. For customer public web research, competition research, industry information collection, this kind of tool chain is the key.

3.3 TUI Official Preview

DeerFlow project comes with Terminal Workbench preview:

! DeerFlow TUI

Pre-sales meaning: It not only has a Web experience, but also supports terminal native workflow, suitable for R & D, data, operation and maintenance, and Agent engineering teams.

3.4 Star History

README built-in Star History diagram:

! Star History Chart

Pre-sales meaning: high community attention can be used as auxiliary evidence of open source maturity and ecological heat, but it cannot be equated with production stability.

Schematic 3.5 Backend Architecture

The core architecture given by the project back-end README can be summarized as follows:

Key points:

-'/api/langgraph/*' is forwarded by Nginx to the Gateway LangGraph-compatible API.

-Gateway assume both REST API and Agent Runtime.

-Lead Agent combines context, upload files, sandbox, memory, skills, children, and clarify processes through the middleware chain.

4. Core Competence List

5. Applicable Scenario

5.1 Presales and Solutions Team

DeerFlow is a great pre-sales research assistant ":

-Enter the customer name, industry, official website and public information.

-Automatically complete the sorting of customer background, business pain points, competition trends and industry trends.

-Output pre-sales call outline, PoC suggestion, scheme structure and demonstration speech.

Advantages:

-Pre-sales methodology can be made into skills.

-The results can be written to the file system to form deliverable materials.

-The use threshold can be lowered through the Web UI or IM portal.

5.2 Enterprise Knowledge Work Automation

Suitable for departments:

-Market research

-Strategic analysis

-Consultant

-Legal/compliance screening

-Product operation

-Content Team

Typical tasks:

-First Draft Industry Report

-Summary of policy impact

-Competition database

-Customer data collation

-Multi-document analysis

-Web page/report/PPT generation

5.3 R & D and technical team

Suitable for tasks:

-Code Repository Understanding

-Small function implementation

-Bug Localization

-Generate scripts and tools

-Code documentation and architecture description

-Access the development process via TUI or API

The project supports Codex CLI, Claude Code OAuth, ACP agent, MCP Server, Docker sandbox, which will be more attractive to technical teams.

5.4 Enterprise Agent Platform PoC

If customers want to build an internal Agent platform instead of just making a single robot, DeerFlow can be used as a reference base:

-Unified model configuration

-Unified tool configuration

-Unified skill library

-Unified sandbox execution

-Unified Web / API / IM portal

-Unified trace and memory management

5.5 multi-terminal enterprise assistant

DeerFlow support Telegram, Slack, Feishu/Lark, Enterprise WeChat, DingTalk, WeChat and other IM channels. For enterprises that already rely heavily on collaborative tools, this means that agents can be embedded in existing work portals rather than opening up a new system.

6. Not quite the scene

7. Architecture, Deployment and Integration

7.1 full stack architecture

DeerFlow is a complete application, not just a library:

7.2 deployment method

Docker is officially recommended, especially for long-term running or shared environments.

7.3 Model Integration

DeerFlow model configuration is managed through 'config.yaml. Support:

-OpenAI

-Anthropic

-DeepSeek

-Gemini

-OpenRouter

-vLLM

-Xiaomi MiMo

-Codex CLI

-Claude Code OAuth

-Arbitrary LangChain-compatible provider

Example:

models:
  - name: gpt-5-responses
    display_name: GPT-5 (Responses API)
    use: langchain_openai:ChatOpenAI
    model: gpt-5
    api_key: $OPENAI_API_KEY
    use_responses_api: true
    output_version: responses/v1

Pre-sales focus: It can be adapted to the customer's existing model gateway, privatized vLLM, OpenAI-compatible services, or combined with volcano engine/BytePlus related models and InfoQuest tools.

7.4 Tools and MCP Integration

DeerFlow support:

-web_search

-web_fetch

-image_search

-file read/write

-bash

-MCP Server

-Python functions

-Custom skills

This provides an entry point for enterprise systems integration, such:

-Internal knowledge base

-CRM

-Work Order System

-Database

-BI Platform

-Documentation system

-Private search

-Code Warehouse

How to use #8.

8.1 Docker Recommended Path

git clone https://github.com/bytedance/deer-flow.git
cd deer-flow

make setup
make docker-init
make docker-start

Access:

http://localhost:2026

8.2 Local Development Path

git clone https://github.com/bytedance/deer-flow.git
cd deer-flow

make config
make check
make install
make dev

8.3 TUI usage

uv pip install 'deerflow-harness[tui]'

deerflow
deerflow --continue
deerflow --resume THREAD
deerflow --print "summarize this repo"
deerflow --json "hello"

Note: the package named "deerflow-harness" on PyPI is displayed as' 0.0.1 'placeholder when queried on 2026-06-30, while the' backend/packages/harness/pyproject.toml' label version' 2.1.0' in the warehouse '. Therefore, the actual use should be based on the warehouse README/source code installation method to avoid misinstallation of placeholder packages.

8.4 Embedded Python Client

from deerflow.client import DeerFlowClient

client = DeerFlowClient()

response = client.chat("Analyze this paper for me", thread_id="my-thread")

for event in client.stream("hello"):
    if event.type == "messages-tuple" and event.data.get("type") == "ai":
        print(event.data["content"])

9. What can I say before sales

9.1 a word

DeerFlow is an open source Super Agent Harness that can upgrade large models from "answering questions" to "completing tasks": it has sandboxes, file systems, skills, memory, child generation, and multi-terminal portals, and is suitable for building enterprise-wide research assistants, content assistants, R & D assistants, and long-task automation platforms.

9.2 for Business Owner

It can help the team to automatically complete repetitive knowledge work such as data collection, research and analysis, report generation, PPT/web page first draft, content production, etc., liberating manual work from "finding data and sorting out formats" and focusing more on judgment and decision-making.

9.3 for Technical Leader

Based on LangGraph and LangChain, the DeerFlow provides complete agent runtime, middleware chain, sandbox, MCP, skills, memory, and subagents. It is not a closed SaaS, but MIT open source, self-hosted, scalable, can be used as an internal Agent platform for engineering reference.

9.4 for security responsible person

The DeerFlow has high permissions such as system command execution, file reading and writing, and business logic invocation, so it cannot expose the public network. The correct methods are local trusted network, strong authentication, IP whitelist, network isolation, sandbox isolation, least privilege, and audit trail.

10. Customer pain point mapping

11. Frequently Asked Customer Questions

12. PoC Recommendations

12.1 PoC 1: Pre-sales Research Assistant

Objective: Enter customer name, official website, industry and competition to automatically generate pre-sales research report.

Recommended Validation:

12.2 PoC II: Enterprise Document Analysis and Report Generation

Objective: Upload multiple PDF/Word/PPT/Excel to generate summary, risk points, recommendations and final report.

Validation Focus:

-File upload and conversion capabilities

-Long context processing capability

-Structured quality of results

-Output file deliverability

-Sensitive data isolation and permissions

12.3 PoC III: R & D Code Assistant

Objective: Let DeerFlow analyze a real-world code repository, complete small fixes or generate architectural notes.

Validation Focus:

-Warehouse understanding ability

-File reading and writing ability

-bash/test execution capability

-TUI experience

-Generate change notes and audit records

12.4 PoC IV: IM Ingress Enterprise Assistant

Objective: To access DeerFlow in Feishu/Enterprise WeChat/DingTalk so that business personnel can submit tasks through chat.

Validation Focus:

-Channel access costs

-User Experience

-Permissions and identity mapping

-File upload/download

-Multi-round task status management

13. Risks and Considerations

13.1 security risk is the first priority

The official README clearly indicates that the DeerFlow has high permission capabilities such as system command execution, resource operation, and business logic invocation, and is designed as a local trusted environment by default. Strict security measures must be taken for public or untrusted network deployment.

Recommendation:

-Don't bare the public.

-Use strong authentication and reverse proxy.

-Configure the IP whitelist.

-The network is isolated to a private VLAN.

-The Docker/Kubernetes sandbox is preferred for the production environment.

-Minimum permissions and audits for high-risk tools.

-Do not allow the Agent to directly operate on the production database.

13.2 Docker Socket Risk

The configuration document points out that some Docker-outside-of-Docker modes need to mount the host Docker socket, and the Docker socket is equivalent to the root-level control capability of the host. Multi-tenant or public network environments should prioritize provisioner/Kubernetes isolation.

Stability of 13.3 2.0 after rewriting

DeerFlow 2.0 is a ground-up rewrite, it means that the architecture has been upgraded, but it also means that:

-Versions may change quickly.

-Documentation and implementation may have update discrepancies.

-Locked version required for production landing.

-Complete regression testing is required.

13.4 PyPI Package Name Note

2026-06-30 Query deerflow-harness PyPI is displayed as placeholder 0.0.1, while the harness version in the warehouse is 2.1.0. The pre-sales demo and PoC should not directly rely on the PyPI placeholder package and should be executed according to the official installation path of the warehouse.

13.5 Cost and Performance

Long tasks, subagents, searches, crawls, code execution, and multiple model calls introduce costs and delays.

PoC is recommended when recording:

-Single task time-consuming

-Number of model calls

-Token consumption

-Number of tool invocations

-Number of failed retries

-Number of manual interventions

-Generate results availability

14. Competition and Alternative Perspectives

Pre-sales judgment: If the customer wants to do "Enterprise Agent Platform PoC", the DeerFlow will produce results faster than the pure framework. If the customer only needs a simple question and answer, it may be too heavy. If the customer needs strict production process control, it should be combined with the workflow/approval/permission system instead of letting the Agent bear all the responsibilities alone.

15. My Pre-Sales Judgment

DeerFlow is an open source project that is well suited as a pre-sales demo and PoC base. Its highlight is not a single point algorithm, but a complete set of engineering capabilities required for Agent landing: Web UI, Gateway, TUI, IM, sandbox, file system, skills, memory, subagents, MCP and tracing.

It is suitable for showing customers the new stage of "AI from chat to task execution", especially suitable for in-depth research, pre-sales data preparation, content generation, document analysis, research and development assistance and enterprise internal Agent platform construction.

But it's also not a silver bullet produced directly out of the box. When landing, security boundaries, authentication and authorization, sandbox isolation, tool permissions, cost control, observability, evaluation system and version governance must be designed simultaneously. It is recommended to use it to make a PoC with clear boundaries before sales: select non-production data, read-only systems, limited tools, and clear output indicators, so that customers can see the business value of long-task Agent first, and then discuss production transformation.

16. REFERENCE

-GitHub-bytedance/deer-flow

-DeerFlow official website

-README Chinese Version

-Install.md

-Backend README

-Configuration Guide

-TUI Guide

-Security Policy

-Release v2.0.0

-MIT License