deer-flow: Orchestrate Long-Running AI Agent Tasks

Overview

DeerFlow is a super-agent harness for coordinating AI work that takes more than a single prompt-response cycle. A lead agent can delegate to subagents and use skills, tools, memory, and sandboxes to research, code, and create across longer tasks.

It is intended for developers and teams who want to run and extend an agent system, rather than adopt a narrowly scoped research library. The repository describes version 2.0 as a ground-up rewrite; the earlier framework remains on the main-1.x branch.

Key Features

• Coordinates a lead agent with subagents for delegated work.
• Provides sandbox modes for local, Docker, and Kubernetes-backed execution.
• Extends agent capabilities through skills and configurable MCP servers.
• Includes short-term context management and long-term memory features.
• Offers web search and crawling integrations, with providers configured by the operator.
• Supports a web interface, messaging channels, and a terminal workbench.
• Provides Docker and local development startup paths, plus SQLite or PostgreSQL for persistent deployments.

Use Cases

• Developers can build research workflows that gather and synthesize information across extended runs.
• Engineering teams can delegate coding and repository tasks to agents with sandboxed execution.
• Operators can offer a shared agent workspace with configurable models, skills, and tool integrations.
• Teams can connect messaging channels so users can submit agent tasks through supported chat platforms.

Project Facts

• Language: Python
• License: MIT
• Stars: 83.4k
• Forks: 11.6k
• Topics: agent, agentic, agentic-framework, agentic-workflow, ai, ai-agents, deep-research, harness, langchain, langgraph, langmanus, llm, multi-agent, nodejs, podcast, python, superagent, typescript
• Archived: No

Getting Started

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

The setup wizard configures an LLM provider and optional integrations. See the README for prerequisites, deployment options, configuration, and security guidance.

Alternatives

• deepagents: Deep Agents provides planning, delegation, file access, and context management in a Python harness, while DeerFlow emphasizes self-hosted, long-running workflows with sandboxes and memory.
• mcp-agent: mcp-agent centers on MCP server integrations and composable workflows, including Temporal execution, rather than DeerFlow’s broader self-hosted agent runtime.
• OpenHuman: OpenHuman packages persistent agents as desktop, browser, and terminal apps as well as a library, while DeerFlow focuses on a self-hosted harness for extended tasks.
• fast-agent: fast-agent emphasizes declarative multimodal agents and MCP workflows, while DeerFlow targets long-running tasks coordinated through subagents, memory, and sandboxes.

Considerations

• Running the system requires configuring an LLM provider, and optional search or tool integrations may require separate credentials.
• The README lists Python 3.12+ and Node.js 22+ for the project, and recommends Docker for deployment.
• Sandboxing and host access depend on configuration. Review the security guidance before exposing a deployment or enabling execution capabilities.
• Version 2.0 is a ground-up rewrite, so users of the original framework should consult the main-1.x branch for that codebase.
• The repository has 895 open issues, so evaluate current maintenance needs and compatibility for your intended deployment.